	A	B	C	D	E	F	G	H	I	J
1	Project ID	Project Name	Project Summary	Team Members	Innovation Expert	Supervisor	Project URL	Contact Email
2	Student ID	Name	Surname

3	T2601	GridLens	Because electricity cannot be stored economically at scale, supply and demand on the grid must be balanced at every moment. Producers therefore commit in advance to how much they will generate the next day. When actual output deviates from this commitment, the system covers the shortfall or surplus with other sources, usually at a higher cost, and this imbalance cost comes directly out of the producer’s profit. For producers whose output depends on uncontrollable factors such as weather, this cost largely depends on the quality of their forecasts. To reduce the risk, producers rely on multiple forecast providers, yet no single forecaster is best under all conditions, and the choice between them is still left to habit or a simple average. GridLens is a forecast referee that fills this gap. By learning how each forecaster errs under different conditions, it predicts tomorrow’s errors in advance and combines the forecasts to minimize the expected cost under market rules. The goal is not the most accurate forecast, but the one expected to cost the least. One of the referee’s forecasters is our own model, which is updated throughout the day as new information arrives, so when something unexpected happens, the referee revises its decision for the remaining hours based on the most current information. Developed on real data from the Turkish electricity market, GridLens will be evaluated not only by forecast accuracy but also by the savings it delivers to producers.	22203816	Ahmet Fatih	Öztürk	Gökalp Cinbiş	Doruk Öner	https://grid-lens-three.vercel.app/	fatihozturk@ug.bilkent.edu.tr
4	22202662	Ece	Bulut
5	22202399	Izzet Yagızefe	Sungurlu
6	22203732	Berke ˙Ismail Erhan	Asıl
7	22102137	Yunus	Tan
8	T2602	DiaGraph	DiaGraph unifies a person’s fragmented health data and reasons over it causally, for both patients and doctors. Records are scattered across hospitals, cities, countries and years, so DiaGraph connects through open standards: FHIR-based patient access APIs, national health records such as Turkey’s e-Nabız, clinical notes in several languages, and wearables. A language model extracts findings from these records, and a calibrated classifier checks each one: whether it is present or negated, whether it refers to the patient, and when it happened. Each finding is mapped to international vocabularies (ICD-10, ICD-11, SNOMED CT) and linked to a medical knowledge graph built from curated sources such as PrimeKG and HPO, never from AI-generated facts. Every edge carries its provenance, and each patient becomes a temporal subgraph. A causal layer separates causal from merely associative relations. On top of this, a team of LLM agents works directly on the graph database through typed tools. Their exact roles are still being designed, but together they will rank differential diagnoses by counterfactual inference over the causal network, using both which findings are present and the order in which they appeared, and suggest the most informative next test. They will also track trends in lab and wearable data and grow the graph from new evidence. Any proposed edge is checked against its evidence and approved by a clinician before it is committed. Patients see their own subgraph in plain language, with a source for every sentence, and are guided to their doctor; they never see ranked diagnoses or high-risk links. Clinicians see the full reasoning only with time-limited consent enforced at the database level. Using synthetic data in more than one language, we evaluate extraction accuracy, diagnostic accuracy against an LLM-only baseline, and the precision of agent-proposed causal edges	22302901	Elif Su	Temirel	Can Koyuncu	Cevdet Aykanat	https://diagraph-site.vercel.app	su.temirel@ug.bilkent.edu.tr, bora.balci@ug.bilkent.edu.tr, asya.unal@ug.bilkent.edu.tr, bilgehan.akin@ug.bilkent.edu.tr, ali.sen@ug.bilkent.edu.tr
9	22301839	Bora	Balcı
10	22204017	Asya	Ünal
11	22203368	Ali	Şen
12	22203657	Bilgehan	Akın
13	T2603	ArtiGlasses	This project aims to develop an automated visual inspection system for detecting defects in product packaging wrappers on a production line using smart glasses, such as Meta Ray-Ban, equipped with image-capturing capabilities. The system will capture images of the package from multiple angles as it is rotated by a human operator or, in future implementations, by a robotic system, ensuring comprehensive coverage of the wrapper surface. The captured images will be transmitted to a computer-based vision system, where image processing and artificial intelligence techniques will be used to identify, classify, and assess potential wrapper defects, such as tears, wrinkles, misalignment, surface damage, and sealing irregularities.	22202350	Hasan Mert	İNCESESLİ	Doruk ÖNER	Doruk ÖNER	https://artiglasses.com/	anil.keskin@ug.bilkent.edu.tr, berhan.kiyanus@ug.bilkent.edu.tr, mert.incesesli@ug.bilkent.edu.tr, emre.bayrak@ug.bilkent.edu.tr, huseyin.aydin@ug.bilkent.edu.tr,
14	22201915	Anıl	Keskin
15	22203742	Berhan	KIYANUS
16	22303713	Kerem Emre	BAYRAK
17	22203683	Muhammed Hüseyin	AYDIN
18	T2604	TRACED	TRACED is an analyst web application that investigates claims shared on Turkish social media. It automatically identifies factual claims in X (Twitter) posts, groups similar statements, and prioritises them based on how quickly they spread. The system retrieves relevant evidence, ranks it using a model fine-tuned for Turkish claims, and produces clear assessments supported by source links and explanations. It also visualises how claims spread and highlights patterns that may indicate coordinated sharing. By bringing claim discovery, evidence-based assessment, and propagation analysis into one workflow, TRACED helps analysts decide what needs attention and make informed, reviewable judgments.	22303060	Emir	Akay	Mesut Gözütok	Özgür Ulusoy	https://traced-bilkent.github.io	mehmetoztur@ug.bilkent.edu.tr
19	22303670	Ateş	Tülümen
20	22301772	Mehmet	Öztürk
21	22302490	Ozan Süha	Atamer
22	22302081	Mustafa	Akseli
23	T2605	SENTINEL	Today, pension fund oversight relies on static rules: they catch yesterday’s violations, but miss emerging risks and raise false alarms whenever the whole market moves. Sentinel changes that. It learns what “normal” looks like for each fund, both against its own history and against its peers, and separates genuine fund-specific anomalies from market-wide noise. Every signal is condensed into a single, explainable risk score, so auditors see not only which funds need attention, but why. Standing watch beside the expert, Sentinel never replaces human judgment; it sharpens it, helping auditors focus on the risks that matter and make faster, better-grounded decision.	222301879	Ayşe Tayyibe Zehra	Özkan	Abdullah Sönmez	Selim Aksoy	https://traced-bilkent.github.io	mailto:tayyibe.ozkan@ug.bilkent.edu.tr
24	22201729	Melike Feyza	Bahadırlı
25	22102837	İlayda	Bozan
26	22203553	Efe	Erdoğmuş
27
28	T2606	VEIL	Generative AI music models learn from large collections of songs that are often scraped without the artists' consent. Once trained, they can imitate an artist's sound, style and even voice on demand. VEIL gives musicians a way to protect their work before releasing it. Our model adds a carefully computed perturbation that listeners cannot hear, but that stops AI models from learning the protected music during training. Generative models trained on protected tracks fail to reproduce the artist and produce distorted, unusable outputs instead. Unlike existing approaches, VEIL separates each song into its components (vocals, drums, bass and other instruments) and shapes the protection for each one according to human hearing, so both instrumental music and singing voice are covered. The protection is also designed to survive common removal attempts such as MP3 compression, denoising and stem separation. The project will deliver an application where users upload a track to protect. The model will then protect the track and the user will receive back the protected track as well as a report that measures how protected the track is and how audible the change is, using objective metrics.	22302122	Bartu	Turan	Haluk Altunel	Sinem Sav	https://www.musicveil.com/seniorproject	TBE
29	22302069	Başak	Ünal
30	22301672	Fatma Ilgın	Vural
31	22302165	Okay Arda	Hamamcı
32	22303187	Selen	Güneri
33	T2607	MONA	Camera application specifically dedicated to cryptographically sign photos the moment they are taken. The signature is then used to verify that photo is from a real, three dimensional scene rather than of a screen, a print, or a model built to deceive the camera. Some photo editing features like blurring, cropping are supported within the system. The edits made can be verified to be bound to some photo that is also signed using the app. Verification process for the validity of a picture edited discloses no additional knowledge.	22303525	Eren	Özilgili	Abdullah Talayhan	Abdullah Talayhan	https://vouch-for-mona.github.io/Mona/	eren.ozilgili@ug.bilkent.edu.tr nur.kabakbas@ug.bilkent.edu.tr berkay.demircin@ug.bilkent.edu.tr erkam.uysal@ug.bilkent.edu.tr berra.alkis@ug.bilkent.edu.tr
34	22302855	Bera Nur	Kabakbaş
35	22303482	Berkay	Demirçin
36	22302205	Erkam	Uysal
37	22302527	Berra	Alkış
38	T2608	ChronoFS	ChronoFS is a filesystem with version control built into its core. The system records changes to files, building a history of the file’s evolution. Because versioning happens at the filesystem level, ChronoFS is designed to capture these changes efficiently, with low overhead and compact storage. A timeline view makes it easy to explore past versions, see how a file has evolved, and restore any earlier state. The system also uses machine learning models for semantic search across different file types and their version histories. These models allow ChronoFS to interpret, query, and compare the content of documents throughout its timeline, enabling more meaningful search and retrieval across a file’s history.	22301699	Ceyhun Deniz	Keleş	Abdullah Talayhan	Abdullah Talayhan	https://chronofs.com	deniz.keles@ug.bilkent.edu.tr
39	22302230	Melis Ece	Aktaş
40	22301780	Doruk	Kolçak
41	22301665	Ahmed Utku	Özüdoğru
42	22302436	Ömer Can	Baykara
43	T2609	AssetForge	This project is an AI-assisted tool designed to help small game studios and independent developers extract fully editable 3D models from a single photograph, ready for direct export to engines like Unity and Godot. It addresses the limitations of current generative AI tools, which typically produce rigid, single-piece meshes that are difficult to modify. By adopting a flexible, component-based geometric approach (initially inspired by methods like 3-Sweep) and integrating modern AI capabilities such as boundary segmentation, texture inpainting, and depth estimation, the tool allows users to maintain control over individual parts. This enables developers to easily create clean, game-ready 3D assets from concept art or reference photos without requiring high-budget design processes or advanced modeling skills.	22201809	Defne Lal	Oruncak	Cem Çimenbiçer	Yusuf Sahillioğlu	https://assetforge.fyrkn.com	lal.oruncak@ug.bilkent.edu.tr
44	22202030	Leyla	Özpınar
45	22103263	Can	Akpınar
46	22102577	Fevzi Yarkın	Müftüoğlu
47
48	T2610	RNAtlas	RNAtlas is an AI-assisted platform for discovering and designing RNA molecules that can bind to a desired biological target. Given a protein, DNA, or RNA target, RNAtlas combines existing scientific knowledge and computational analysis to identify known RNA binders or propose new candidates when suitable ones are unavailable. It then evaluates these candidates and provides an interactive 3D representation of their interactions with the target. Additionally, by identifying weak regions of an interaction, RNAtlas can further improve the RNA design. Overall, the platform aims to provide an integrated workflow that guides users from a biological target to promising, computationally evaluated RNA-binding candidates.	22203840	Mehmet Batuhan	Neziroğlu	A. Ercüment Çiçek	A. Ercüment Çiçek	https://rnatlas.github.io/	tugce.neziroglu@ug.bilkent.edu.tr
49	22203839	Elif Tuğçe	Neziroğlu
50	22201748	Alperen	Karslıoğlu
51	22203676	Furkan	Karademir
52	22301777	Ahmet Eren	Gözübenli
53	T2611	MapTheGap	Our project is an AI-powered mobile road monitoring and driver assistance system that assesses and maps overall road quality, including potholes, uneven surfaces, and other road irregularities, without requiring dedicated hardware. Users can contribute road data dynamically as they drive using a smartphone camera or a compatible vehicle camera. The system combines road images, smartphone accelerometer readings, GPS, and detected sudden-braking events to identify potential road defects and hazards, estimate their severity where possible, and associate observations with geographic locations. Instead of continuously processing or storing video, the application adapts its image analysis frequency based on motion events and driving conditions to reduce computational and battery overhead. These observations build an evolving road-quality map, letting the system identify newly reported issues, update existing records as more vehicles pass through an area, and filter out unreliable observations by comparing multiple reports and data sources. Users who do not contribute camera images can still access previously collected road information and receive location-based audio warnings as they approach reported hazards or poor-quality road sections. Beyond assisting individual drivers, the collected information could help road authorities, such as the General Directorate of Highways (KGM), a potential institutional customer, monitor road conditions and prioritize maintenance. By combining AI-based visual analysis, motion and braking data, GPS, and crowdsourced observations, the system aims to provide continuously updated road-quality information that supports safer driving and road maintenance planning.	22302223	Eylül İdil	Yıldız	Cem Çimenbiçer	Ayşegül Dündar Boral	https://mapthegap.vercel.app/	efe.mutlu@ug.bilkent.edu.tr
54	22301742	Eser Tekin	Tekeli
55	22303326	Mehmet Efe	Mutlu
56	22302862	Hikmet Şamil	Karaş
57	22202803	Muhammed Devran	Yılmaz
58	T2612	Upseen	Our project aims to develop a platform that measures and optimizes brand visibility across generative AI search engines such as ChatGPT, Gemini, and Perplexity. As AI-powered search transforms how consumers discover and evaluate brands, businesses need new ways to understand and strengthen their presence in AI-generated responses. Our platform will track brand mentions, recommendations, and competitive positioning to provide measurable visibility insights and actionable Generative Engine Optimization (GEO) strategies. By combining automated visibility monitoring, competitive analysis, and optimization recommendations, the project aims to help businesses improve their discoverability and competitiveness in the evolving AI search ecosystem.	22203720	Hüseyin Deniz	Kurtulan	Harun Ünlüsoy	Fazlı Can	https://upseen.co/	TBE
59	22201699	Mert Okhan	İbrahimağaoğlu
60	22201808	Burakhan	Kaldırım
61	22102554	Ferit	Bilgi
62	22003413	Abdullah Batuhan	Yorulmaz
63	T2613	WereVul	WereVul is an LLM-assisted framework that automatically detects, localizes, and repairs security vulnerabilities in Java repositories. Manually auditing large codebases is slow, static analysis tools drown developers in false positives, and AI models can produce convincing but wrong answers. WereVul addresses this by combining static analysis with locally running code language models: it narrows a repository down to the relevant files and functions, identifies vulnerabilities down to the line level, classifies them by CWE category, and proposes patches. No AI-generated fix is trusted blindly; every patch is validated through compilation, proof-of-vulnerability, and regression tests. Because everything runs locally, source code never leaves the developer’s environment. The result is clear, actionable reports that help small teams, AppSec engineers, and open-source maintainers secure their projects without dedicated security expertise.	22203033	Turan Emre	Adıgüzel	Onur Ayhan	Anıl Koyuncu	https://werevul.github.io/	ovunc.senol@ug.bilkent.edu.tr zeynep.tasevli@ug.bilkent.edu.tr eduman@ug.bilkent.edu.tr merve.topuz@ug.bilkent.edu.tr emre.adiguzel@ug.bilkent.edu.tr
64	22303597	Elif	Duman
65	22303819	Umut Övünç	Şenol
66	22201936	Zeynep	Taşevli
67	22302937	Merve	Topuz
68	T2614	PAWN	PAWN (Photogrammetric Autonomous Workspace Navigator) turns an off-the-shelf, sub-100-gram DJI Tello into an autonomous indoor search-and-rescue scout. The drone has no onboard computer, depth sensor, LiDAR or GPS: its only output is a monocular 2D video stream and its only input is four-axis control commands. All intelligence runs on a ground station, which uses visual SLAM to build a 3D digital twin of a building the drone has never seen before, explores it in real time while reacting to moving obstacles, and detects a designated target such as a person. When the target is found, PAWN marks its location and the return path in the 3D map, delivering what a skilled human pilot would, from a cheap, safe, palm-sized drone.	22201710	Burkay	Tunçtürk	Ali Cengiz Beğen	Salih Özgür Öğüz	https://pawn-air.github.io/PAWN	burkay.tuncturk@ug.bilkent.edu.tr
69	22202567	Doğa	Ertürk
70	22301926	Lütfü Heysem	Kızıloğlu
71	22301195	Nijat	Mukhtarov
72	22203465	Şebnem Bihter	Keleş
73	T2615	ZeroTrust	What if you could collaborate with zero infrastructure? No Internet, no servers, no 3rd parties. ZeroTrust is a local, peer-to-peer, encrypted communication and collaboration platform that trusts no one, not even your own router. Whether you handle sensitive data, want better privacy, or simple don't have the infrastructure, our project can connect you to your peers, with no middleman. Already have a network? Our project can create a secure collaboration platform right on top of it, with no unwelcome guests. Heading somewhere remote? Any router or even your phone will do, regardless of Internet access. A natural fit for defence industry, emergency responders, air-gapped facilities, hospitals, critical-infrastructure operators and more.	22201022	Rasul	Ibrahimzade	Mohammad Mahmoody	Mohammad Mahmoody	https://cs491-2026-zerotrust.github.io/CS491-ZeroTrust/	rasul.ibrahimzade@ug.bilkent.edu.tr
74	22302443	Ali	Demir
75	22201366	Ansar	Myrzagerey
76	22303555	Mehmet İlbey	Angur
77	22301544	Amir Mousa Hasan	Othman
78	T2616	Mendrey	Fixing a software bug today is a slow, manual relay between support, testers and developers. Our project replaces it with an end-to-end multi-agent system that takes a bug from the user's first report to a verified, review-ready patch in a single automated pipeline. Specialized AI agents each own one step of the process: refining the report, reproducing the bug, locating its cause in the codebase, generating a minimal fix, and independently reviewing that fix. A central orchestrator coordinates the agents and enforces clear rules on what each one is allowed to do, bringing in a human engineer whenever a decision is risky or uncertain. Because the system is tailored to its own codebase, tests and development practices, it delivers faster and more reliable bug fixing.	22203140	Atakan	Akar	Yunus Emre Uluçay	Eray Tüzün	https://mendrey.com/	Serdar.kara@ug.bilkent.edu.tr
79	22303215	Arda Ege	Erdoğan
80	22203391	Erdem	Uğurlu
81	22202442	Kıvanç Kuzey	Dikici
82	22202880	Serdar	Kara
83	T2617	SUST-I	Our project aims to develop an AI-powered mobile application that helps users make sustainable decisions about products they no longer use, promoting a circular economy. By analyzing a product's image, condition, and other relevant information, the application recommends suitable end-of-use strategies, such as reuse, repair, repurposing, or recycling, and explains the reasoning behind each recommendation. These recommendations are connected to a location-based system that helps users discover nearby repair services, recycling points, and people who can reuse, share, or lend items. Users can also express their needs, allowing the system to proactively match them with nearby opportunities. To encourage long-term engagement, the application incorporates gamification and a sustainability feed featuring practical tips, DIY ideas, news, and local events. By combining end-of-use management with community-driven resource sharing, our project aims to extend product lifecycles, reduce waste, and make sustainable living more accessible and actionable.	22301304	Nabeeha	Khan	Tuğrul Ünsal	Cevdet Aykanat	https://sust-i.online/	sudenaz.yilmaz@ug.bilkent.edu.tr mehmet.bars@ug.bilkent.edu.tr vildan.cetin@ug.bilkent.edu.tr nabeeha.khan@ug.bilkent.edu.tr
84	22001875	Mehmet	Bars
85	22103486	Sudenaz	Yılmaz
86	22203353	Ayşe Vildan	Çetin
87
88	T2618	Minos	Companies are rapidly deploying LLM-based tools such as chatbots, code review bots, and test generators, yet they lack a reliable way to know whether these tools actually work, either before release or once users depend on them. Minos is an evaluation platform into which a company can plug any LLMpowered tool with minimal setup. It automatically builds an evaluation pipeline by mining the tool’s existing artifacts, scores outputs with LLM judges that are continuously calibrated against feedback from the company’s own employees, and reports how much each verdict can be trusted. Minos acts as a quality gate in the CI/CD pipeline and keeps monitoring the tool in production, so regressions are caught by the team rather than by users. We are building and validating Minos together with a large industry partner, starting with a customer-facing chatbot.	22202039	Veli	Karakaya	Tunga Bayrak	Eray Tüzün	https://minoseval.com/	yigit.korkmaz@ug.bilkent.edu.tr
89	22201749	Yusuf Yiğit	Korkmaz
90	22202876	Alp Eren	Köken
91	22201904	Selman	Şahin
92	22201659	Barış	Peksak
93	T2619	DOST	DOST (Detect, Observe, Shield, Transfer) is a local, non-bypassable security layer that keeps sensitive company data out of LLMs and other AI tools. Powered by an on-device model that tells sensitive from safe content, it catches confidential information before it leaves the machine and lets only sanitized data through, so teams can use AI freely without risking leaks.	22302136	Ahmet Eren	Gökalp	Alper Sarıkan	Ayşegül Dündar	https://dost-ai-privacy.github.io/DOST-AI-Privacy/	eren.gokalp@ug.bilkent.edu.tr
94	22301791	Can	Yıldırım
95	22301923	Ali Emir	Alsırt
96	22301884	Atahan	Alp
97	22302877	Ahmet Hakan	Kızılçubuk
98	T2620	VETeran	VETeran is an AI-powered veterinary decision support platform designed to help pet owners make informed decisions about their pets' health. By combining visual symptom analysis with intelligent clinical triage, VETeran aims to provide reliable, evidence-based guidance on the urgency of veterinary care. The platform will leverage Multimodal Retrieval-Augmented Generation to analyze pet photos against trusted veterinary cases, alongside a hybrid retrieval system that integrates clinical safety rules with relevant medical literature. This approach prioritizes pet safety while ensuring that every recommendation is transparent, explainable, and supported by credible sources. Through rigorous testing and veterinary expert validation, VETeran aims to bridge the gap between uncertainty and professional veterinary care, empowering pet owners to take timely, appropriate action without replacing the expertise of veterinarians.	22303229	Doruk	Arca	Deniz Katırcıoğlu-Öztürk	Mohammad Mahmoody	https://veteran2027.github.io/	veteran_team@googlegroups.com
99	22302973	İsmail Yağız	Güven
100	22302048	Kaan Yusuf	Kayhan
101	22302782	Erez	Koç
102	22302269	Nevzat Eren	Tezcan
103	T2621	PRISM	PRISM (Pull Request Intelligence for Software Maintenance) is a code review assistant built on a deterministic CIA core. When a PR is created or updated, PRISM combines call graph-based dependency analysis with repository history, such as code churn, bug frequency, and co-change information, to identify the software entities that may be affected by the change and to calculate a risk score. These results reach the reviewer in several forms: a structured impact report; visualizations such as an impact graph and, where applicable, class or sequence diagrams, supported by a behavioral summary of affected tests and API endpoints; and an AI review that uses the deterministic results as context.	22202103	Emine İrem	Esendemir	Mert Cemri	Eray Tüzün	https://prismai-devtools.github.io/	irem.esendemir@ug.bilkent.edu.tr
104	22202709	Halil Arda	Özongun
105	22203822	Sude	Ergün
106	22202444	Işıl	Ünveren
107	22202829	Yağmur	Demirer
108	T2622	PixelMind	Mobile game studios still rely heavily on manual QA because automated options fall short: random-tap tools do not understand the game and get stuck, while AI agents that send every frame to large multimodal models are too expensive to run at scale. PixelMind is an autonomous agent that tests 2D mobile games like a human tester, at a fraction of that cost. Lightweight local vision models turn each screen into a structured description of buttons, text and game state; a small decision layer handles routine actions, and a large language model is called only when the agent gets stuck. PixelMind explores games on its own, reports freezes, soft-locks and rendering glitches with steps to reproduce, and will be benchmarked on bug-detection rate and cost per test hour against existing approaches.	22201738	Deniz	Çağlar	Enis Kerem Çakmak	Selim Aksoy	https://pixelmind-bilkent.github.io/PixelMind	engin.ermis@ug.bilkent.edu.tr
109	22202639	Eda Eylül	Aslan
110	22203167	Engin	Ermiş
111	22201745	Erenalp	Yuca
112	22202788	Kerem	Varnalı
113	T2623	Revivant	An interactive 3D reconstruction and visualization system can digitally preserve buildings affected by earthquakes and other disasters. Using photographs and video recordings captured before or after a disaster, the system will create navigable virtual environments that allow users to explore documented spaces, examine visible damage, and compare building conditions over time. By combining 3D Gaussian Splatting with interactive visualization and evidence-based documentation, the system aims to support disaster assessment, engineering research, and the preservation of places with historical or personal significance. Particular emphasis will be placed on reconstruction reliability, clearly identifying missing or uncertain information, and protecting the privacy of individuals whose photographs and personal spaces are represented.	22302644	Muhammet Furkan	Demir	Ateş Akaydın	Uğur Güdükbay	https://revivant-t2623.github.io/todo	dilara.erbenzer@ug.bilkent.edu.tr
114	22301802	Furkan	Kırat
115	22303774	Dilara	Erbenzer
116	22302913	Gökay	Nuray
117	22303210	Berke	Halavurt
118	T2624	Movewise	Moving to a new home raises two questions: how should it be moved and how should it look? Movewise answers both in one system. Users scan their current and new homes with a phone, and Movewise turns these scans into measured 3D models. On these models, users design the layout of their new home and confirm that every item fits, both in its new room and along the way through doors, corridors, and stairs. From the final design, Movewise creates a step-by-step moving plan for removal, truck loading, and placement. Individuals see their new home designed before moving day, and moving companies get a reliable estimate of the crew and time each job needs.	22201530	Mehmet	Dizdar	Veysi İşler	Uğur Güdükbay	https://movewise-project.github.io/	mehmet.dizdar@ug.bilkent.edu.tr
119	22203236	Hüsnü	Arkca
120	22201566	Aral	Müftüoğlu
121	22202950	Onur Ege	Yakar
122	22202610	Yunus Emre	Aslan
123	T2625	SENV	SENV is a modular runtime security gateway designed to protect enterprise AI agent workflows from indirect prompt injection, privileged tool misuse, and data exfiltration without altering underlying foundation models. Sitting as a protective middleware between autonomous agents and enterprise execution environments (databases, APIs, and file systems), SENV intercepts privileged actions before execution. It evaluates proposed tool calls through deterministic capability policies and dual-execution behavioral validation to detect state hijacking and context poisoning in real time. By pairing runtime interception with an automated, graph-based adversarial red-teaming engine, the platform establishes provable runtime safety boundaries for multi-agent enterprise deployments.	22303059	Atilla	Akkuş	Erhan Dolak	Mohammad Mahmoody	https://teamssenv.github.io/SSenv_Proposal/
124	22302593	Atilla	Ündar
125	22301917	Arda	Yüksel
126	22302163	Gökdeniz	Derelioğlu
127	22301064	Muhammad Musa	Qureshi
128	T2626	MINDCROWD	Our project aims to develop a modular, behavior-aware crowd simulation and management plugin for Unity that enables more realistic, intelligent, and efficient interactions among non-player characters (NPCs). Built on Unity's existing NavMesh system, the plugin will enhance traditional navigation by integrating crowd-aware pathfinding, environmental perception, personality-driven behavior, emotional responses, social group coordination, and resource management. Rather than simply following predefined paths, NPCs will dynamically adapt their decisions and movements based on surrounding agents, environmental events, and changing crowd conditions. Through configurable components, developers will be able to selectively enable these capabilities and customize NPC behavior for different games and simulations. By bringing these interconnected systems together in a reusable framework, our project seeks to improve the realism, flexibility, and performance of large-scale crowd simulations while providing developers with intuitive tools to manage and analyze complex crowd behavior.	22203359	Başar	Çubukçuoğlu	Cem Çimenbiçer	Uğur Güdükbay	https://mindcrowd.github.io/webpage/#home	basar.cubukcuoglu@ug.bilkent.edu.tr
129	22202664	Muhammed Emir	Eren
130	22202801	Ata Eren	Kılıç
131	22203670	Yunus Emre	Erkan
132	22203455	Emir	Atlas
133	T2627	WoundTrack	The aim of this project is to develop a digital platform for monitoring chronic wounds, focusing on diabetic foot ulcers and pressure ulcers in home healthcare services. Currently, wound assessment often depends on regular home visits by doctors, who may need to send wound images to specialists for visual evaluation. These images and related patient information are not stored or shared in a standardized and secure way, making it difficult to monitor healing progress and potentially delaying specialist consultation. The platform will guide users in capturing standardized wound images. Machine learning models and image segmentation techniques will be used to measure wound area and identify ulcer stages. The system will combine image findings with clinical data to assess infection risk and monitor changes over time. Since the platform processes sensitive medical data, it will also investigate privacy-preserving machine learning methods, including federated learning and homomorphic encryption, to enable model training and analysis while keeping patient data secure and, where possible, stored locally. The platform will allow doctors, nurses, and other healthcare professionals to securely review relevant patient data, improving coordination and supporting earlier clinical decisions.	22303759	Aybegüm Fatma	Çelebi	Deniz Katırcıoğlu Öztürk	Sinem Sav	https://wound-track-web.vercel.app/
134	22302768	Beren	Irgat
135	22302236	Aslı	Kanber
136	22302184	Nehir	Konaş
137	22202206	Ceyda	Yıldırım
138	T2628	Data Protection System in Fleet Vehicles (Dapsiv)	When a drone is lost in a place with no internet, such as a disaster zone or a contested area, everything it recorded is lost with it. Superstes is a decentralized "black box" for drone swarms that removes that single point of failure. Each drone continuously encodes its flight data using Reed-Solomon erasure coding and spreads the resulting fragments across its peers over an ad-hoc wireless network. Because any sufficiently large subset of fragments is enough to reconstruct the original, a drone's complete record survives even when several swarm members are lost, at a fraction of the storage and radio cost of full replication. The system needs no server, no internet connection and no central coordinator, and its redundancy level is tunable to the mission's risk profile. Our contribution is the design, implementation and evaluation of this resilient storage layer, validated on low-cost ESP32 microcontroller nodes under realistic swarm conditions such as mobility, packet loss and sudden failures.	22302993	Yusuf Bedri	Bitiren	Abdullah Talayhan	Abdullah Talayhan	https://yusuf-bedri-bitiren.github.io/dapsiv/	bedri.bitiren@ug.bilkent.edu.tr
139	22201545	Altan	Altay
140	22203529	Mehmet Akif	Yavuz
141	22301744	İbrahim Enes	Yılmaz
142	22202359	Bedirhan	Omaç
143	T2629	PickWise	Our project aims to simplify the process of analyzing Turkish e-commerce listings in terms of credibility. It can be used by online shoppers to help them avoid risky listings on e-commerce platforms. To achieve this, by analyzing key factors such as seller ratings, review authenticity, product descriptions, images, and price comparisons, our system provides users with a trust score alongside a comprehensive credibility report for any listing they view. By evaluating each parameter individually, we catch details that users might easily overlook and combine these findings, allowing users to assess everything at a single glance.	22203347	Yunus Dildar	Çelik	Alper Çevik	Selim Aksoy	https://yunuscelik21.github.io/PickWise/	ruveyda.bilir@ug.bilkent.edu.tr; dildar.celik@ug.bilkent.edu.tr; berkay.ataman@ug.bilkent.edu.tr; hilal.cosgun@ug.bilkent.edu.tr; toprak.yildizeli@ug.bilkent.edu.tr
144	22301745	Berkay	Ataman
145	22201571	Hilal	Coşgun
146	22203082	Rüveyda	Bilir
147	22303329	Volkan Toprak	Yıldızeli
148	T2630	DiaComp	DiaComp is a calm digital companion designed to reduce the mental burden of living with Type 1 diabetes. By continuously observing glucose data and learning from a user's past experiences, it aims to identify moments when timely support may be useful without adding to the constant stream of alarms and notifications already present in diabetes management. Rather than replacing existing medical devices or treatment plans, DiaComp focuses on the overlooked emotional and cognitive side of the condition: helping users feel less alone, less anxious, and less responsible for watching every fluctuation themselves. Its core innovation is combining deterministic safety rules with personalized historical context and carefully constrained language-model-generated guidance to provide support only when it is genuinely needed.	22203094	Melike	Kara	Haluk Altunel	Salih Özgür Öğüz	https://diacomp-bilkent.github.io	simay.bal@ug.bilkent.edu.tr
149	22302724	Atakan	Kaya
150	22302444	Mert	Suci
151	22202896	Ahmet	Büyükberber
152	22301635	Simay	Bal
153	T2631	Vernier	Vernier is a personal memory layer for AI assistants that keeps up with how users' lives change. Today's memory systems store facts about users but rarely notice when those facts go out of date, so assistants confidently repeat cancelled plans, old preferences, and expired deadlines. Vernier tracks the lifecycle of every stored fact (current, updated, cancelled, or expired), retrieves the most current version, and signals when memory does not hold a reliable answer. Built on our team's published research, it reaches developers as an API and SDK, with a dashboard where users can see, correct, and delete what is remembered about them.	22101494	Ansuman	Mullick	Eray Tüzün	Fazli Can	https://www.dexico.org/landing#research	ansuman.mullick@ug.bilkent.edu.tr
154	22103096	Yağmur	Yılmazer
155	21902094	Barış	Çelik
156	22102378	Arif Emre	Burhan
157	22002553	Osman	Baktır
158	T2632	GENIE5 (GEne Network Inference from Evidence)	Genie5 turns a biomedical research question into a directed, evidence-backed map of gene regulation. Statistical analysis can reveal that two genes are linked, but rarely which one regulates the other, so researchers still settle the question by hand-searching the literature, one gene pair at a time. Genie5 automates that review. It collects evidence from scientific papers and curated biological databases, reconciles conflicting reports, and determines the direction and type of each interaction, such as activation or repression. Researchers receive an interactive 2D/3D network that they can explore, verify and export to their existing tools, including CRNA for analysis on their own expression data.	22302668	Sırrı Kaan	Oğuzkan	Gurkan Bebek	Can Alkan	https://twin-pathways.github.io/	ceylin.atalar@ug.bilkent.edu.tr
159	22302229	Derin Ceylin	Atalar
160	22302289	Ayça	Topkaya
161	22303934	Alper	Koçyiğit
162	22302218	Ömer Turab	Memiş
163	T2633	Sourceline	Embedded devices such as routers and IoT products ship firmware built from dozens of open-source libraries, but vendors rarely say which components and versions are inside. When a new vulnerability is published, device owners and security teams have no reliable way to know whether they are affected. Sourceline answers that question from the firmware image itself. It extracts the filesystem from Linux-based firmware, analyzes the stripped ELF binaries compiled for MIPS and ARM, and identifies the open-source libraries and their exact versions. It does this by matching function-level fingerprints against a reference corpus that we build by cross-compiling many library versions with different compilers, optimization levels and target architectures. The result is a recovered Software Bill of Materials (SBOM), mapped to known CVEs and shown in a web application as a clear vulnerability report. Commercial tools in this space publish no accuracy figures. Sourceline will be evaluated on an open, reproducible benchmark, so its results can be measured and checked by anyone.	22302697	Alperen	Kalyoncu	Muhammed Ali Yurdagül	Abdullah Talayhan	https://sourceline-cs491.github.io	rifat.arifoglu@ug.bilkent.edu.tr
164	22303115	Arda Selim	Arslan
165	22303392	Deniz Can	Yoldaş
166	22303937	Emre	Altuner
167	22303017	Rıfat	Arifoğlu
168	T2634	OmniVision	We are developing an industrial vision platform that transforms operators' monitoring requests in natural language into workflows coordinated by an agentic orchestration layer. This layer dynamically selects and combines reusable skills such as detection, tracking, and event recognition, and manages agents responsible for planning, reviewing, feedback, annotation, and communication between components. As industrial cameras capture footage, vision-language models analyze relevant frames in real time manifested by operator's input. The platform matches task requirements to available capabilities, selects models according to current hardware resources, and coordinates data retrieval, annotation, automated fine-tuning, and validation when capabilities are missing. Operator approval and feedback memory guide execution and adaptation to changing requirements, reducing manual monitoring and custom development effort.	22203024	Altay İlker	Yiğitel	Doruk Öner	Doruk Öner	https://omnivision.cs491.workers.dev/	perhat.amanlyyev@ug.bilkent.edu.tr
169	22202238	Barış	Pamuksuz
170	22203114	Emiralp	İlgen
171	22201895	Mustafa Mert	Gülhan
172	22201007	Perhat	Amanlyyev
173	T2635	BehaviorLens	BehaviorLens is an AI-powered user experience (UX) testing and customer behavior analysis platform designed to simulate, analyze, and understand diverse user behaviors in web environments. The project focuses on developing customizable, domain-specific AI personas whose behavioral characteristics can be adjusted through parameters such as personality traits, digital experience, patience, and risk tolerance. Powered by large language model (LLM) agents, these personas interact with websites, perform predefined tasks, and generate behavioral traces that help businesses and UX researchers examine different user journeys and identify potential usability issues. The platform also aims to support companies in analyzing their customer profiles and generating personalized UX recommendations based on dynamic personas, either defined by the platform or created and refined using customer data provided by businesses. From a research perspective, the project investigates whether different persona parameters produce consistent and measurable behavioral differences and how these differences can contribute to usability evaluation and customer behavior analysis. Ultimately, BehaviorLens aims to provide businesses with an adaptable, data-driven platform for understanding their target audiences, identifying usability challenges, and improving digital experiences through AI-powered behavioral simulations.	22201687	İlayda	Altınok	Haluk Altunel	Anıl Koyuncu	https://behaviorlens.github.io/cs491-project-website/	ilayda.altinok@ug.bilkent.edu.tr; yasemin.altun@ug.bilkent.edu.tr; said.bakan@ug.bilkent.edu.tr; berkay.simsek@ug.bilkent.edu.tr; amina.azizli@ug.bilkent.edu.tr
174	22202739	Yasemin	Altun
175	22201008	Amina	Azizli
176	22303338	Berkay	Şimşek
177	22302852	Emir Said	Bakan
178	T2636	ConsoliData	Our project aims to build a platform that reads and makes sense of scattered data arriving from different departments and companies in various formats, such as Excel, PDF, Word, or scanned documents, without requiring any fixed template, and brings it together into a single, structured data source. We want to automate the collection and cleanup work that is still largely done by hand today, which is slow and error-prone, while keeping every piece of information traceable back to the document it came from. Rather than replacing ERP systems, our solution works alongside them to structure the data they don't cover, and we aim to give organizations with heavy reporting loads, especially multi-subsidiary holding companies and financial institutions, a faster, more reliable, and more auditable reporting process.	22201765	Mehmet Emre	Şahin	Onur Ayhan	Özgür Salih Öğüz	https://kenanataman.github.io/consolidata/	emresahin@ug.bilkent.edu.tr
179	22203434	Ahmet Kenan	Ataman
180	22201637	Ece	Şeşen
181	22203549	Umut Çağın	Özcan
182	22202085	Nevzat Han	Altın
183	T2637	PromptShield	PromptShield is a desktop AI assistant that lets people use powerful cloud AI models without handing over their secrets. Today, users must choose between cloud assistants that are capable but send everything to external servers, including API keys, passwords, client names and private code, and local models that keep data private but are too weak for complex work. PromptShield removes this trade-off. It searches the user's own files and answers each question in one of two modes: in strict mode, a local model answers and nothing leaves the computer; in sanitized mode, a cloud model such as Claude answers, but only after a two-layer privacy gateway has detected and masked any sensitive data. Every outgoing prompt is recorded in an audit log, so users and companies can see exactly what left the machine and what did not.	22301172	Leyla	Abbasova	Alper Sarikan	Cevdet Aykanat	https://leyla2109.github.io/PromptShield/
184	22201178	Zaeem Masood	Sheikh
185	22301186	Farrukh	Mammadov
186	22301574	Shahin	Ibrahimli
187

Project ID

Project Name

Project Summary

Team Members

Innovation Expert

Supervisor

Project URL

Contact Email

Student ID

Name

Surname

T2601

GridLens

Because electricity cannot be stored economically at scale, supply and demand on the grid must be balanced at every moment. Producers therefore commit in advance to how much they will generate the next day. When actual output deviates from this commitment, the system covers the shortfall or surplus with other sources, usually at a higher cost, and this imbalance cost comes directly out of the producer’s profit. For producers whose output depends on uncontrollable factors such as weather, this cost largely depends on the quality of their forecasts. To reduce the risk, producers rely on multiple forecast providers, yet no single forecaster is best under all conditions, and the choice between them is still left to habit or a simple average. GridLens is a forecast referee that fills this gap. By learning how each forecaster errs under different conditions, it predicts tomorrow’s errors in advance and combines the forecasts to minimize the expected cost under market rules. The goal is not the most accurate forecast, but the one expected to cost the least. One of the referee’s forecasters is our own model, which is updated throughout the day as new information arrives, so when something unexpected happens, the referee revises its decision for the remaining hours based on the most current information. Developed on real data from the Turkish electricity market, GridLens will be evaluated not only by forecast accuracy but also by the savings it delivers to producers.

22203816

Ahmet Fatih

Öztürk

Gökalp Cinbiş

Doruk Öner

https://grid-lens-three.vercel.app/

fatihozturk@ug.bilkent.edu.tr

22202662

Ece

Bulut

22202399

Izzet Yagızefe

Sungurlu

22203732

Berke ˙Ismail Erhan

Asıl

22102137

Yunus

Tan

T2602

DiaGraph

DiaGraph unifies a person’s fragmented health data and reasons over it causally, for both patients and doctors. Records are scattered across hospitals, cities, countries and years, so DiaGraph connects through open standards: FHIR-based patient access APIs, national health records such as Turkey’s e-Nabız, clinical notes in several languages, and wearables. A language model extracts findings from these records, and a calibrated classifier checks each one: whether it is present or negated, whether it refers to the patient, and when it happened. Each finding is mapped to international vocabularies (ICD-10, ICD-11, SNOMED CT) and linked to a medical knowledge graph built from curated sources such as PrimeKG and HPO, never from AI-generated facts. Every edge carries its provenance, and each patient becomes a temporal subgraph. A causal layer separates causal from merely associative relations. On top of this, a team of LLM agents works directly on the graph database through typed tools. Their exact roles are still being designed, but together they will rank differential diagnoses by counterfactual inference over the causal network, using both which findings are present and the order in which they appeared, and suggest the most informative next test. They will also track trends in lab and wearable data and grow the graph from new evidence. Any proposed edge is checked against its evidence and approved by a clinician before it is committed. Patients see their own subgraph in plain language, with a source for every sentence, and are guided to their doctor; they never see ranked diagnoses or high-risk links. Clinicians see the full reasoning only with time-limited consent enforced at the database level. Using synthetic data in more than one language, we evaluate extraction accuracy, diagnostic accuracy against an LLM-only baseline, and the precision of agent-proposed causal edges

22302901

Elif Su

Temirel

Can Koyuncu

Cevdet Aykanat

https://diagraph-site.vercel.app

su.temirel@ug.bilkent.edu.tr, bora.balci@ug.bilkent.edu.tr, asya.unal@ug.bilkent.edu.tr, bilgehan.akin@ug.bilkent.edu.tr, ali.sen@ug.bilkent.edu.tr

22301839

Bora

Balcı

22204017

Asya

Ünal

22203368

Ali

Şen

22203657

Bilgehan

Akın

T2603

ArtiGlasses

This project aims to develop an automated visual inspection system for detecting defects in product packaging wrappers on a production line using smart glasses, such as Meta Ray-Ban, equipped with image-capturing capabilities. The system will capture images of the package from multiple angles as it is rotated by a human operator or, in future implementations, by a robotic system, ensuring comprehensive coverage of the wrapper surface. The captured images will be transmitted to a computer-based vision system, where image processing and artificial intelligence techniques will be used to identify, classify, and assess potential wrapper defects, such as tears, wrinkles, misalignment, surface damage, and sealing irregularities.

22202350

Hasan Mert

İNCESESLİ

Doruk ÖNER

https://artiglasses.com/

anil.keskin@ug.bilkent.edu.tr, berhan.kiyanus@ug.bilkent.edu.tr, mert.incesesli@ug.bilkent.edu.tr, emre.bayrak@ug.bilkent.edu.tr, huseyin.aydin@ug.bilkent.edu.tr,

22201915

Anıl

Keskin

22203742

Berhan

KIYANUS

22303713

Kerem Emre

BAYRAK

22203683

Muhammed Hüseyin

AYDIN

T2604

TRACED

TRACED is an analyst web application that investigates claims shared on Turkish social media. It automatically identifies factual claims in X (Twitter) posts, groups similar statements, and prioritises them based on how quickly they spread. The system retrieves relevant evidence, ranks it using a model fine-tuned for Turkish claims, and produces clear assessments supported by source links and explanations. It also visualises how claims spread and highlights patterns that may indicate coordinated sharing. By bringing claim discovery, evidence-based assessment, and propagation analysis into one workflow, TRACED helps analysts decide what needs attention and make informed, reviewable judgments.

22303060

Emir

Akay

Mesut Gözütok

Özgür Ulusoy

https://traced-bilkent.github.io

mehmetoztur@ug.bilkent.edu.tr

22303670

Ateş

Tülümen

22301772

Mehmet

Öztürk

22302490

Ozan Süha

Atamer

22302081

Mustafa

Akseli

T2605

SENTINEL

Today, pension fund oversight relies on static rules: they catch yesterday’s violations, but miss emerging risks and raise false alarms whenever the whole market moves. Sentinel changes that. It learns what “normal” looks like for each fund, both against its own history and against its peers, and separates genuine fund-specific anomalies from market-wide noise. Every signal is condensed into a single, explainable risk score, so auditors see not only which funds need attention, but why. Standing watch beside the expert, Sentinel never replaces human judgment; it sharpens it, helping auditors focus on the risks that matter and make faster, better-grounded decision.

222301879

Ayşe Tayyibe Zehra

Özkan

Abdullah Sönmez

Selim Aksoy

https://traced-bilkent.github.io

mailto:tayyibe.ozkan@ug.bilkent.edu.tr

22201729

Melike Feyza

Bahadırlı

22102837

İlayda

Bozan

22203553

Efe

Erdoğmuş

T2606

VEIL

Generative AI music models learn from large collections of songs that are often scraped without the artists' consent. Once trained, they can imitate an artist's sound, style and even voice on demand. VEIL gives musicians a way to protect their work before releasing it. Our model adds a carefully computed perturbation that listeners cannot hear, but that stops AI models from learning the protected music during training. Generative models trained on protected tracks fail to reproduce the artist and produce distorted, unusable outputs instead. Unlike existing approaches, VEIL separates each song into its components (vocals, drums, bass and other instruments) and shapes the protection for each one according to human hearing, so both instrumental music and singing voice are covered. The protection is also designed to survive common removal attempts such as MP3 compression, denoising and stem separation. The project will deliver an application where users upload a track to protect. The model will then protect the track and the user will receive back the protected track as well as a report that measures how protected the track is and how audible the change is, using objective metrics.

22302122

Bartu

Turan

Haluk Altunel

Sinem Sav

https://www.musicveil.com/seniorproject

TBE

22302069

Başak

Ünal

22301672

Fatma Ilgın

Vural

22302165

Okay Arda

Hamamcı

22303187

Selen

Güneri

T2607

MONA

Camera application specifically dedicated to cryptographically sign photos the moment they
are taken. The signature is then used to verify that photo is from a real, three dimensional
scene rather than of a screen, a print, or a model built to deceive the camera. Some photo
editing features like blurring, cropping are supported within the system. The edits made can
be verified to be bound to some photo that is also signed using the app. Verification process
for the validity of a picture edited discloses no additional knowledge.

22303525

Eren

Özilgili

Abdullah Talayhan

https://vouch-for-mona.github.io/Mona/

eren.ozilgili@ug.bilkent.edu.tr nur.kabakbas@ug.bilkent.edu.tr berkay.demircin@ug.bilkent.edu.tr erkam.uysal@ug.bilkent.edu.tr berra.alkis@ug.bilkent.edu.tr

22302855

Bera Nur

Kabakbaş

22303482

Berkay

Demirçin

22302205

Erkam

Uysal

22302527

Berra

Alkış

T2608

ChronoFS

ChronoFS is a filesystem with version control built into its core. The system records
changes to files, building a history of the file’s evolution. Because versioning happens at the
filesystem level, ChronoFS is designed to capture these changes efficiently, with low overhead
and compact storage. A timeline view makes it easy to explore past versions, see how a file
has evolved, and restore any earlier state. The system also uses machine learning models for
semantic search across different file types and their version histories. These models allow
ChronoFS to interpret, query, and compare the content of documents throughout its timeline,
enabling more meaningful search and retrieval across a file’s history.

22301699

Ceyhun Deniz

Keleş

Abdullah Talayhan

https://chronofs.com

deniz.keles@ug.bilkent.edu.tr

22302230

Melis Ece

Aktaş

22301780

Doruk

Kolçak

22301665

Ahmed Utku

Özüdoğru

22302436

Ömer Can

Baykara

T2609

AssetForge

This project is an AI-assisted tool designed to help small game studios and independent
developers extract fully editable 3D models from a single photograph, ready for direct export
to engines like Unity and Godot. It addresses the limitations of current generative AI tools,
which typically produce rigid, single-piece meshes that are difficult to modify. By adopting a
flexible, component-based geometric approach (initially inspired by methods like 3-Sweep)
and integrating modern AI capabilities such as boundary segmentation, texture inpainting, and
depth estimation, the tool allows users to maintain control over individual parts. This enables
developers to easily create clean, game-ready 3D assets from concept art or reference photos
without requiring high-budget design processes or advanced modeling skills.

22201809

Defne Lal

Oruncak

Cem Çimenbiçer

Yusuf Sahillioğlu

https://assetforge.fyrkn.com

lal.oruncak@ug.bilkent.edu.tr

22202030

Leyla

Özpınar

22103263

Can

Akpınar

22102577

Fevzi Yarkın

Müftüoğlu

T2610

RNAtlas

RNAtlas is an AI-assisted platform for discovering and designing RNA molecules that can bind to a desired
biological target. Given a protein, DNA, or RNA target, RNAtlas combines existing scientific knowledge and
computational analysis to identify known RNA binders or propose new candidates when suitable ones are
unavailable. It then evaluates these candidates and provides an interactive 3D representation of their interactions
with the target. Additionally, by identifying weak regions of an interaction, RNAtlas can further improve the
RNA design. Overall, the platform aims to provide an integrated workflow that guides users from a biological
target to promising, computationally evaluated RNA-binding candidates.

22203840

Mehmet Batuhan

Neziroğlu

A. Ercüment Çiçek

https://rnatlas.github.io/

tugce.neziroglu@ug.bilkent.edu.tr

22203839

Elif Tuğçe

Neziroğlu

22201748

Alperen

Karslıoğlu

22203676

Furkan

Karademir

22301777

Ahmet Eren

Gözübenli

T2611

MapTheGap

Our project is an AI-powered mobile road monitoring and driver assistance system that assesses
and maps overall road quality, including potholes, uneven surfaces, and other road irregularities,
without requiring dedicated hardware. Users can contribute road data dynamically as they drive
using a smartphone camera or a compatible vehicle camera. The system combines road images,
smartphone accelerometer readings, GPS, and detected sudden-braking events to identify potential
road defects and hazards, estimate their severity where possible, and associate observations with
geographic locations. Instead of continuously processing or storing video, the application adapts its
image analysis frequency based on motion events and driving conditions to reduce computational
and battery overhead. These observations build an evolving road-quality map, letting the system
identify newly reported issues, update existing records as more vehicles pass through an area, and
filter out unreliable observations by comparing multiple reports and data sources. Users who do not
contribute camera images can still access previously collected road information and receive
location-based audio warnings as they approach reported hazards or poor-quality road sections.
Beyond assisting individual drivers, the collected information could help road authorities, such as
the General Directorate of Highways (KGM), a potential institutional customer, monitor road
conditions and prioritize maintenance. By combining AI-based visual analysis, motion and braking
data, GPS, and crowdsourced observations, the system aims to provide continuously updated
road-quality information that supports safer driving and road maintenance planning.

22302223

Eylül İdil

Yıldız

Cem Çimenbiçer

Ayşegül Dündar Boral

https://mapthegap.vercel.app/

efe.mutlu@ug.bilkent.edu.tr

22301742

Eser Tekin

Tekeli

22303326

Mehmet Efe

Mutlu

22302862

Hikmet Şamil

Karaş

22202803

Muhammed Devran

Yılmaz

T2612

Upseen

Our project aims to develop a platform that measures and optimizes brand visibility across
generative AI search engines such as ChatGPT, Gemini, and Perplexity. As AI-powered search
transforms how consumers discover and evaluate brands, businesses need new ways to
understand and strengthen their presence in AI-generated responses. Our platform will track
brand mentions, recommendations, and competitive positioning to provide measurable visibility
insights and actionable Generative Engine Optimization (GEO) strategies. By combining
automated visibility monitoring, competitive analysis, and optimization recommendations, the
project aims to help businesses improve their discoverability and competitiveness in the
evolving AI search ecosystem.

22203720

Hüseyin Deniz

Kurtulan

Harun Ünlüsoy

Fazlı Can

https://upseen.co/

TBE

22201699

Mert Okhan

İbrahimağaoğlu

22201808

Burakhan

Kaldırım

22102554

Ferit

Bilgi

22003413

Abdullah Batuhan

Yorulmaz

T2613

WereVul

WereVul is an LLM-assisted framework that automatically detects, localizes, and repairs
security vulnerabilities in Java repositories. Manually auditing large codebases is slow,
static analysis tools drown developers in false positives, and AI models can produce
convincing but wrong answers. WereVul addresses this by combining static analysis with
locally running code language models: it narrows a repository down to the relevant files
and functions, identifies vulnerabilities down to the line level, classifies them by CWE
category, and proposes patches. No AI-generated fix is trusted blindly; every patch is
validated through compilation, proof-of-vulnerability, and regression tests. Because
everything runs locally, source code never leaves the developer’s environment. The result
is clear, actionable reports that help small teams, AppSec engineers, and open-source
maintainers secure their projects without dedicated security expertise.

22203033

Turan Emre

Adıgüzel

Onur Ayhan

Anıl Koyuncu

https://werevul.github.io/

ovunc.senol@ug.bilkent.edu.tr zeynep.tasevli@ug.bilkent.edu.tr eduman@ug.bilkent.edu.tr merve.topuz@ug.bilkent.edu.tr emre.adiguzel@ug.bilkent.edu.tr

22303597

Elif

Duman

22303819

Umut Övünç

Şenol

22201936

Zeynep

Taşevli

22302937

Merve

Topuz

T2614

PAWN

PAWN (Photogrammetric Autonomous Workspace Navigator) turns an off-the-shelf, sub-100-gram DJI Tello into an
autonomous indoor search-and-rescue scout. The drone has no onboard computer, depth sensor, LiDAR or GPS: its only
output is a monocular 2D video stream and its only input is four-axis control commands. All intelligence runs on a
ground station, which uses visual SLAM to build a 3D digital twin of a building the drone has never seen before, explores
it in real time while reacting to moving obstacles, and detects a designated target such as a person. When the target is
found, PAWN marks its location and the return path in the 3D map, delivering what a skilled human pilot would, from a
cheap, safe, palm-sized drone.

22201710

Burkay

Tunçtürk

Ali Cengiz Beğen

Salih Özgür Öğüz

https://pawn-air.github.io/PAWN

burkay.tuncturk@ug.bilkent.edu.tr

22202567

Doğa

Ertürk

22301926

Lütfü Heysem

Kızıloğlu

22301195

Nijat

Mukhtarov

22203465

Şebnem Bihter

Keleş

T2615

ZeroTrust

What if you could collaborate with zero infrastructure? No Internet, no servers, no 3rd parties.
ZeroTrust is a local, peer-to-peer, encrypted communication and collaboration platform that trusts no
one, not even your own router. Whether you handle sensitive data, want better privacy, or simple don't
have the infrastructure, our project can connect you to your peers, with no middleman. Already have a
network? Our project can create a secure collaboration platform right on top of it, with no unwelcome
guests. Heading somewhere remote? Any router or even your phone will do, regardless of Internet
access. A natural fit for defence industry, emergency responders, air-gapped facilities, hospitals,
critical-infrastructure operators and more.

22201022

Rasul

Ibrahimzade

Mohammad Mahmoody

https://cs491-2026-zerotrust.github.io/CS491-ZeroTrust/

rasul.ibrahimzade@ug.bilkent.edu.tr

22302443

Ali

Demir

22201366

Ansar

Myrzagerey

22303555

Mehmet İlbey

Angur

22301544

Amir Mousa Hasan

Othman

T2616

Mendrey

Fixing a software bug today is a slow, manual relay between support, testers and developers. Our project
replaces it with an end-to-end multi-agent system that takes a bug from the user's first report to a verified,
review-ready patch in a single automated pipeline. Specialized AI agents each own one step of the
process: refining the report, reproducing the bug, locating its cause in the codebase, generating a minimal
fix, and independently reviewing that fix. A central orchestrator coordinates the agents and enforces clear
rules on what each one is allowed to do, bringing in a human engineer whenever a decision is risky or
uncertain. Because the system is tailored to its own codebase, tests and development practices, it delivers
faster and more reliable bug fixing.

22203140

Atakan

Akar

Yunus Emre Uluçay

Eray Tüzün

https://mendrey.com/

Serdar.kara@ug.bilkent.edu.tr

22303215

Arda Ege

Erdoğan

22203391

Erdem

Uğurlu

22202442

Kıvanç Kuzey

Dikici

22202880

Serdar

Kara

T2617

SUST-I

Our project aims to develop an AI-powered mobile application that helps users make sustainable
decisions about products they no longer use, promoting a circular economy. By analyzing a product's
image, condition, and other relevant information, the application recommends suitable end-of-use
strategies, such as reuse, repair, repurposing, or recycling, and explains the reasoning behind each
recommendation. These recommendations are connected to a location-based system that helps users
discover nearby repair services, recycling points, and people who can reuse, share, or lend items. Users
can also express their needs, allowing the system to proactively match them with nearby opportunities. To
encourage long-term engagement, the application incorporates gamification and a sustainability feed
featuring practical tips, DIY ideas, news, and local events. By combining end-of-use management with
community-driven resource sharing, our project aims to extend product lifecycles, reduce waste, and
make sustainable living more accessible and actionable.

22301304

Nabeeha

Khan

Tuğrul Ünsal

Cevdet Aykanat

https://sust-i.online/

sudenaz.yilmaz@ug.bilkent.edu.tr mehmet.bars@ug.bilkent.edu.tr vildan.cetin@ug.bilkent.edu.tr nabeeha.khan@ug.bilkent.edu.tr

22001875

Mehmet

Bars

22103486

Sudenaz

Yılmaz

22203353

Ayşe Vildan

Çetin

T2618

Minos

Companies are rapidly deploying LLM-based tools such as chatbots, code review bots, and test
generators, yet they lack a reliable way to know whether these tools actually work, either before release
or once users depend on them. Minos is an evaluation platform into which a company can plug any LLMpowered tool with minimal setup. It automatically builds an evaluation pipeline by mining the tool’s
existing artifacts, scores outputs with LLM judges that are continuously calibrated against feedback from
the company’s own employees, and reports how much each verdict can be trusted. Minos acts as a
quality gate in the CI/CD pipeline and keeps monitoring the tool in production, so regressions are caught
by the team rather than by users. We are building and validating Minos together with a large industry
partner, starting with a customer-facing chatbot.

22202039

Veli

Karakaya

Tunga Bayrak

Eray Tüzün

https://minoseval.com/

yigit.korkmaz@ug.bilkent.edu.tr

22201749

Yusuf Yiğit

Korkmaz

22202876

Alp Eren

Köken

22201904

Selman

Şahin

22201659

Barış

Peksak

T2619

DOST

DOST (Detect, Observe, Shield, Transfer) is a local, non-bypassable security layer that keeps sensitive company data out of LLMs and other AI tools. Powered by an on-device model that tells sensitive from safe content, it catches confidential information before it leaves the machine and lets only sanitized data through, so teams can use AI freely without risking leaks.

22302136

Ahmet Eren

Gökalp

Alper Sarıkan

Ayşegül Dündar

https://dost-ai-privacy.github.io/DOST-AI-Privacy/

eren.gokalp@ug.bilkent.edu.tr

22301791

Can

Yıldırım

22301923

Ali Emir

Alsırt

22301884

Atahan

Alp

22302877

Ahmet Hakan

Kızılçubuk

T2620

VETeran

VETeran is an AI-powered veterinary decision support platform designed to help pet owners make informed decisions about their pets' health. By combining visual symptom analysis with intelligent clinical triage, VETeran aims to provide reliable, evidence-based guidance on the urgency of veterinary care. The platform will leverage Multimodal Retrieval-Augmented Generation to analyze pet photos against trusted veterinary cases, alongside a hybrid retrieval system that integrates clinical safety rules with relevant medical literature. This approach prioritizes pet safety while ensuring that every recommendation is transparent, explainable, and supported by credible sources. Through rigorous testing and veterinary expert validation, VETeran aims to bridge the gap between uncertainty and professional veterinary care, empowering pet owners to take timely, appropriate action without replacing the expertise of veterinarians.

22303229

Doruk

Arca

Deniz Katırcıoğlu-Öztürk

Mohammad Mahmoody

https://veteran2027.github.io/

veteran_team@googlegroups.com

22302973

İsmail Yağız

Güven

100

22302048

Kaan Yusuf

Kayhan

101

22302782

Erez

Koç

102

22302269

Nevzat Eren

Tezcan

103

T2621

PRISM

PRISM (Pull Request Intelligence for Software Maintenance) is a code review assistant built on a deterministic CIA core. When a PR is created or updated, PRISM combines call graph-based dependency analysis with repository history, such as code churn, bug frequency, and co-change information, to identify the software entities that may be affected by the change and to calculate a risk score. These results reach the reviewer in several forms: a structured impact report; visualizations such as an impact graph and, where applicable, class or sequence diagrams, supported by a behavioral summary of affected tests and API endpoints; and an AI review that uses the deterministic results as context.

22202103

Emine İrem

Esendemir

Mert Cemri

Eray Tüzün

https://prismai-devtools.github.io/

irem.esendemir@ug.bilkent.edu.tr

104

22202709

Halil Arda

Özongun

105

22203822

Sude

Ergün

106

22202444

Işıl

Ünveren

107

22202829

Yağmur

Demirer

108

T2622

PixelMind

Mobile game studios still rely heavily on manual QA because automated options fall short: random-tap
tools do not understand the game and get stuck, while AI agents that send every frame to large multimodal
models are too expensive to run at scale. PixelMind is an autonomous agent that tests 2D mobile games
like a human tester, at a fraction of that cost. Lightweight local vision models turn each screen into a
structured description of buttons, text and game state; a small decision layer handles routine actions, and
a large language model is called only when the agent gets stuck. PixelMind explores games on its own,
reports freezes, soft-locks and rendering glitches with steps to reproduce, and will be benchmarked on
bug-detection rate and cost per test hour against existing approaches.

22201738

Deniz

Çağlar

Enis Kerem Çakmak

Selim Aksoy

https://pixelmind-bilkent.github.io/PixelMind

engin.ermis@ug.bilkent.edu.tr

109

22202639

Eda Eylül

Aslan

110

22203167

Engin

Ermiş

111

22201745

Erenalp

Yuca

112

22202788

Kerem

Varnalı

113

T2623

Revivant

An interactive 3D reconstruction and visualization system can digitally preserve buildings affected
by earthquakes and other disasters. Using photographs and video recordings captured before or after
a disaster, the system will create navigable virtual environments that allow users to explore
documented spaces, examine visible damage, and compare building conditions over time.
By combining 3D Gaussian Splatting with interactive visualization and evidence-based documentation,
the system aims to support disaster assessment, engineering research, and the preservation of places
with historical or personal significance. Particular emphasis will be placed on reconstruction reliability,
clearly identifying missing or uncertain information, and protecting the privacy of individuals whose
photographs and personal spaces are represented.

22302644

Muhammet Furkan

Demir

Ateş Akaydın

Uğur Güdükbay

https://revivant-t2623.github.io/todo

dilara.erbenzer@ug.bilkent.edu.tr

114

22301802

Furkan

Kırat

115

22303774

Dilara

Erbenzer

116

22302913

Gökay

Nuray

117

22303210

Berke

Halavurt

118

T2624

Movewise

Moving to a new home raises two questions: how should it be moved and how
should it look? Movewise answers both in one system. Users scan their current and
new homes with a phone, and Movewise turns these scans into measured 3D
models. On these models, users design the layout of their new home and confirm
that every item fits, both in its new room and along the way through doors,
corridors, and stairs. From the final design, Movewise creates a step-by-step
moving plan for removal, truck loading, and placement. Individuals see their new
home designed before moving day, and moving companies get a reliable estimate
of the crew and time each job needs.

22201530

Mehmet

Dizdar

Veysi İşler

Uğur Güdükbay

https://movewise-project.github.io/

mehmet.dizdar@ug.bilkent.edu.tr

119

22203236

Hüsnü

Arkca

120

22201566

Aral

Müftüoğlu

121

22202950

Onur Ege

Yakar

122

22202610

Yunus Emre

Aslan

123

T2625

SENV

SENV is a modular runtime security gateway designed to protect enterprise AI agent workflows from indirect prompt injection, privileged tool misuse, and data exfiltration without altering underlying foundation models. Sitting as a protective middleware between autonomous agents and enterprise execution environments (databases, APIs, and file systems), SENV intercepts privileged actions before execution. It evaluates proposed tool calls through deterministic capability policies and dual-execution behavioral validation to detect state hijacking and context poisoning in real time. By pairing runtime interception with an automated, graph-based adversarial red-teaming engine, the platform establishes provable runtime safety boundaries for multi-agent enterprise deployments.

22303059

Atilla

Akkuş

Erhan Dolak

Mohammad Mahmoody

https://teamssenv.github.io/SSenv_Proposal/

124

22302593

Atilla

Ündar

125

22301917

Arda

Yüksel

126

22302163

Gökdeniz

Derelioğlu

127

22301064

Muhammad Musa

Qureshi

128

T2626

MINDCROWD

Our project aims to develop a modular, behavior-aware crowd simulation and management plugin for Unity that enables more realistic, intelligent, and efficient interactions among non-player characters (NPCs). Built on Unity's existing NavMesh system, the plugin will enhance traditional navigation by integrating crowd-aware pathfinding, environmental perception, personality-driven behavior, emotional responses, social group coordination, and resource management. Rather than simply following predefined paths, NPCs will dynamically adapt their decisions and movements based on surrounding agents, environmental events, and changing crowd conditions. Through configurable components, developers will be able to selectively enable these capabilities and customize NPC behavior for different games and simulations. By bringing these interconnected systems together in a reusable framework, our project seeks to improve the realism, flexibility, and performance of large-scale crowd simulations while providing developers with intuitive tools to manage and analyze complex crowd behavior.

22203359

Başar

Çubukçuoğlu

Cem Çimenbiçer

Uğur Güdükbay

https://mindcrowd.github.io/webpage/#home

basar.cubukcuoglu@ug.bilkent.edu.tr

129

22202664

Muhammed Emir

Eren

130

22202801

Ata Eren

Kılıç

131

22203670

Yunus Emre

Erkan

132

22203455

Emir

Atlas

133

T2627

WoundTrack

The aim of this project is to develop a digital platform for monitoring chronic wounds, focusing on diabetic foot ulcers and pressure ulcers in home healthcare services. Currently, wound assessment often depends on regular home visits by doctors, who may need to send wound images to specialists for visual evaluation. These images and related patient information are not stored or shared in a standardized and secure way, making it difficult to monitor healing progress and potentially delaying specialist consultation. The platform will guide users in capturing standardized wound images. Machine learning models and image segmentation techniques will be used to measure wound area and identify ulcer stages. The system will combine image findings with clinical data to assess infection risk and monitor changes over time. Since the platform processes sensitive medical data, it will also investigate privacy-preserving machine learning methods, including federated learning and homomorphic encryption, to enable model training and analysis while keeping patient data secure and, where possible, stored locally. The platform will allow doctors, nurses, and other healthcare professionals to securely review relevant patient data, improving coordination and supporting earlier clinical decisions.

22303759

Aybegüm Fatma

Çelebi

Deniz Katırcıoğlu Öztürk

Sinem Sav

https://wound-track-web.vercel.app/

134

22302768

Beren

Irgat

135

22302236

Aslı

Kanber

136

22302184

Nehir

Konaş

137

22202206

Ceyda

Yıldırım

138

T2628

Data Protection System in Fleet Vehicles (Dapsiv)

When a drone is lost in a place with no internet, such as a disaster zone or a contested area, everything it recorded is lost with it. Superstes is a decentralized "black box" for drone swarms that removes that single point of failure. Each drone continuously encodes its flight data using Reed-Solomon erasure coding and spreads the resulting fragments across its peers over an ad-hoc wireless network. Because any sufficiently large subset of fragments is enough to reconstruct the original, a drone's complete record survives even when several swarm members are lost, at a fraction of the storage and radio cost of full replication. The system needs no server, no internet connection and no central coordinator, and its redundancy level is tunable to the mission's risk profile. Our contribution is the design, implementation and evaluation of this resilient storage layer, validated on low-cost ESP32 microcontroller nodes under realistic swarm conditions such as mobility, packet loss and sudden failures.

22302993

Yusuf Bedri

Bitiren

Abdullah Talayhan

https://yusuf-bedri-bitiren.github.io/dapsiv/

bedri.bitiren@ug.bilkent.edu.tr

139

22201545

Altan

Altay

140

22203529

Mehmet Akif

Yavuz

141

22301744

İbrahim Enes

Yılmaz

142

22202359

Bedirhan

Omaç

143

T2629

PickWise

Our project aims to simplify the process of analyzing Turkish e-commerce listings in terms of credibility. It can be used by online shoppers to help them avoid risky listings on e-commerce platforms. To achieve this, by analyzing key factors such as seller ratings, review authenticity, product descriptions, images, and price comparisons, our system provides users with a trust score alongside a comprehensive credibility report for any listing they view. By evaluating each parameter individually, we catch details that users might easily overlook and combine these findings, allowing users to assess everything at a single glance.

22203347

Yunus Dildar

Çelik

Alper Çevik

Selim Aksoy

https://yunuscelik21.github.io/PickWise/

ruveyda.bilir@ug.bilkent.edu.tr; dildar.celik@ug.bilkent.edu.tr; berkay.ataman@ug.bilkent.edu.tr; hilal.cosgun@ug.bilkent.edu.tr; toprak.yildizeli@ug.bilkent.edu.tr

144

22301745

Berkay

Ataman

145

22201571

Hilal

Coşgun

146

22203082

Rüveyda

Bilir

147

22303329

Volkan Toprak

Yıldızeli

148

T2630

DiaComp

DiaComp is a calm digital companion designed to reduce the mental burden of living with Type 1 diabetes. By continuously observing glucose data and learning from a user's past experiences, it aims to identify moments when timely support may be useful without adding to the constant stream of alarms and notifications already present in diabetes management. Rather than replacing existing medical devices or treatment plans, DiaComp focuses on the overlooked emotional and cognitive side of the condition: helping users feel less alone, less anxious, and less responsible for watching every fluctuation themselves. Its core innovation is combining deterministic safety rules with personalized historical context and carefully constrained language-model-generated guidance to provide support only when it is genuinely needed.

22203094

Melike

Kara

Haluk Altunel

Salih Özgür Öğüz

https://diacomp-bilkent.github.io

simay.bal@ug.bilkent.edu.tr

149

22302724

Atakan

Kaya

150

22302444

Mert

Suci

151

22202896

Ahmet

Büyükberber

152

22301635

Simay

Bal

153

T2631

Vernier

Vernier is a personal memory layer for AI assistants that keeps up with how users' lives change. Today's memory systems store facts about users but rarely notice when those facts go out of date, so assistants confidently repeat cancelled plans, old preferences, and expired deadlines. Vernier tracks the lifecycle of every stored fact (current, updated, cancelled, or expired), retrieves the most current version, and signals when memory does not hold a reliable answer. Built on our team's published research, it reaches developers as an API and SDK, with a dashboard where users can see, correct, and delete what is remembered about them.

22101494

Ansuman

Mullick

Eray Tüzün

Fazli Can

https://www.dexico.org/landing#research

ansuman.mullick@ug.bilkent.edu.tr

154

22103096

Yağmur

Yılmazer

155

21902094

Barış

Çelik

156

22102378

Arif Emre

Burhan

157

22002553

Osman

Baktır

158

T2632

GENIE5 (GEne Network Inference from Evidence)

Genie5 turns a biomedical research question into a directed, evidence-backed map of gene regulation. Statistical analysis can reveal that two genes are linked, but rarely which one regulates the other, so researchers still settle the question by hand-searching the literature, one gene pair at a time. Genie5 automates that review. It collects evidence from scientific papers and curated biological databases, reconciles conflicting reports, and determines the direction and type of each interaction, such as activation or repression. Researchers receive an interactive 2D/3D network that they can explore, verify and export to their existing tools, including CRNA for analysis on their own expression data.

22302668

Sırrı Kaan

Oğuzkan

Gurkan Bebek

Can Alkan

https://twin-pathways.github.io/

ceylin.atalar@ug.bilkent.edu.tr

159

22302229

Derin Ceylin

Atalar

160

22302289

Ayça

Topkaya

161

22303934

Alper

Koçyiğit

162

22302218

Ömer Turab

Memiş

163

T2633

Sourceline

Embedded devices such as routers and IoT products ship firmware built from dozens of open-source libraries, but vendors rarely say which components and versions are inside. When a new vulnerability is published, device owners and security teams have no reliable way to know whether they are affected. Sourceline answers that question from the firmware image itself. It extracts the filesystem from Linux-based firmware, analyzes the stripped ELF binaries compiled for MIPS and ARM, and identifies the open-source libraries and their exact versions. It does this by matching function-level fingerprints against a reference corpus that we build by cross-compiling many library versions with different compilers, optimization levels and target architectures. The result is a recovered Software Bill of Materials (SBOM), mapped to known CVEs and shown in a web application as a clear vulnerability report. Commercial tools in this space publish no accuracy figures. Sourceline will be evaluated on an open, reproducible benchmark, so its results can be measured and checked by anyone.

22302697

Alperen

Kalyoncu

Muhammed Ali Yurdagül

Abdullah Talayhan

https://sourceline-cs491.github.io

rifat.arifoglu@ug.bilkent.edu.tr

164

22303115

Arda Selim

Arslan

165

22303392

Deniz Can

Yoldaş

166

22303937

Emre

Altuner

167

22303017

Rıfat

Arifoğlu

168

T2634

OmniVision

We are developing an industrial vision platform that transforms operators' monitoring requests in natural language into workflows coordinated by an agentic orchestration layer. This layer dynamically selects and combines reusable skills such as detection, tracking, and event recognition, and manages agents responsible for planning, reviewing, feedback, annotation, and communication between components. As industrial cameras capture footage, vision-language models analyze relevant frames in real time manifested by operator's input. The platform matches task requirements to available capabilities, selects models according to current hardware resources, and coordinates data retrieval, annotation, automated fine-tuning, and validation when capabilities are missing. Operator approval and feedback memory guide execution and adaptation to changing requirements, reducing manual monitoring and custom development effort.

22203024

Altay İlker

Yiğitel

Doruk Öner

https://omnivision.cs491.workers.dev/

perhat.amanlyyev@ug.bilkent.edu.tr

169

22202238

Barış

Pamuksuz

170

22203114

Emiralp

İlgen

171

22201895

Mustafa Mert

Gülhan

172

22201007

Perhat

Amanlyyev

173

T2635

BehaviorLens

BehaviorLens is an AI-powered user experience (UX) testing and customer behavior analysis platform designed to simulate, analyze, and understand diverse user behaviors in web environments. The project focuses on developing customizable, domain-specific AI personas whose behavioral characteristics can be adjusted through parameters such as personality traits, digital experience, patience, and risk tolerance. Powered by large language model (LLM) agents, these personas interact with websites, perform predefined tasks, and generate behavioral traces that help businesses and UX researchers examine different user journeys and identify potential usability issues. The platform also aims to support companies in analyzing their customer profiles and generating personalized UX recommendations based on dynamic personas, either defined by the platform or created and refined using customer data provided by businesses. From a research perspective, the project investigates whether different persona parameters produce consistent and measurable behavioral differences and how these differences can contribute to usability evaluation and customer behavior analysis. Ultimately, BehaviorLens aims to provide businesses with an adaptable, data-driven platform for understanding their target audiences, identifying usability challenges, and improving digital experiences through AI-powered behavioral simulations.

22201687

İlayda

Altınok

Haluk Altunel

Anıl Koyuncu

https://behaviorlens.github.io/cs491-project-website/

ilayda.altinok@ug.bilkent.edu.tr; yasemin.altun@ug.bilkent.edu.tr; said.bakan@ug.bilkent.edu.tr; berkay.simsek@ug.bilkent.edu.tr; amina.azizli@ug.bilkent.edu.tr

174

22202739

Yasemin

Altun

175

22201008

Amina

Azizli

176

22303338

Berkay

Şimşek

177

22302852

Emir Said

Bakan

178

T2636

ConsoliData

Our project aims to build a platform that reads and makes sense of scattered data arriving from different departments and companies in various formats, such as Excel, PDF, Word, or scanned documents, without requiring any fixed template, and brings it together into a single, structured data source. We want to automate the collection and cleanup work that is still largely done by hand today, which is slow and error-prone, while keeping every piece of information traceable back to the document it came from. Rather than replacing ERP systems, our solution works alongside them to structure the data they don't cover, and we aim to give organizations with heavy reporting loads, especially multi-subsidiary holding companies and financial institutions, a faster, more reliable, and more auditable reporting process.

22201765

Mehmet Emre

Şahin

Onur Ayhan

Özgür Salih Öğüz

https://kenanataman.github.io/consolidata/

emresahin@ug.bilkent.edu.tr

179

22203434

Ahmet Kenan

Ataman

180

22201637

Ece

Şeşen

181

22203549

Umut Çağın

Özcan

182

22202085

Nevzat Han

Altın

183

T2637

PromptShield

PromptShield is a desktop AI assistant that lets people use powerful cloud AI models without handing over their secrets. Today, users must choose between cloud assistants that are capable but send everything to external servers, including API keys, passwords, client names and private code, and local models that keep data private but are too weak for complex work. PromptShield removes this trade-off. It searches the user's own files and answers each question in one of two modes: in strict mode, a local model answers and nothing leaves the computer; in sanitized mode, a cloud model such as Claude answers, but only after a two-layer privacy gateway has detected and masked any sensitive data. Every outgoing prompt is recorded in an audit log, so users and companies can see exactly what left the machine and what did not.

22301172

Leyla

Abbasova

Alper Sarikan

Cevdet Aykanat

https://leyla2109.github.io/PromptShield/

184

22201178

Zaeem Masood

Sheikh

185

22301186

Farrukh

Mammadov

186

22301574

Shahin

Ibrahimli

187