I'm a senior (4th-year) Computer Engineering student at Kocaeli University. I work on multilingual NLP and on measuring where models and classifiers actually break — and I build the products those models sit inside. Most of what's below started as a university project and grew into something I kept iterating on.
- 🔬 NLP/ML research at the Kocaeli University Embedded Systems Laboratory — a study I designed and ran there is now a first-author paper in preparation
- 🏢 Previously IT Project Student at BSH Home Appliances (Aug 2025 – Sep 2026) — MuleSoft data integrations and AWS (Lambda, S3, Athena) for the PCM-Integration team
- 🤝 Some of the work below was built with teammates
- 🎻 Away from the keyboard: violin, sketching, and swimming
Zero-shot NACE classification of German trade-register texts — Python · Sentence-Transformers · scikit-learn
A zero-shot classifier is handed the names of its classes and nothing else. Rewriting those names into real definitions is usually treated as prompt-fiddling; it isn't. What a description buys is alignment with the vocabulary of the documents it must attract, and how much that is worth can be measured before accuracy is.
- +26.8 points Top-1 from rewriting the class descriptions as definitions — 81.6% Top-3 on 299 corpus-sampled documents, against 46.8% for an oracle majority-class floor
- Controlled three ways: it is not the language (a literal German rendering of the same content is 2.3 points worse), not the encoder, and not one corpus — on 20 Newsgroups the same work buys nothing, and that exception is the point
- Three label-free studies preregistered — predictions committed to git before any accuracy existed; 12 of 15 held
- A headline number of my own was retracted: 80.0% measured on hand-written examples became 52.5% on real records. The research record above lists every claim the project made and what became of it, including two of its own hypotheses that the evidence destroyed
- Every figure regenerated by
make reproduce; 46 comparisons Holm-corrected; 106 tests and CI
| Project | What it does | Built with |
|---|---|---|
| SenseSign 🤝 (graduation project · private while in development) | Bidirectional Turkish Sign Language communication — my idea and my research design. MediaPipe landmarks feed LSTM / Transformer / ST-GCN recognisers compared on a signer-independent split; an LLM layer turns gloss sequences into Turkish; an avatar path runs the other way. Evaluation is designed up front — modality and window-size ablations, BLEU/chrF/ROUGE-L with human rating and Cohen's κ, McNemar between model pairs. Landmark extraction runs in the browser, so raw video never leaves the client | Python · MediaPipe · FastAPI · React |
| Lumos (live) | An assistant that walks first-time investors through risk profiling and portfolio building, backed by a free-tier LLM chain (Gemini, with Groq and OpenRouter as automatic fallbacks) instead of being locked to one provider. The model never does arithmetic — deterministic engines do, and tests enforce that boundary. Every return is shown real as well as nominal; a user carrying card debt is told to clear it first, above the portfolio rather than below it | FastAPI · React · Gemini/Groq/OpenRouter |
| unshackled-llm-bridge (private, in testing) | A browser extension that moves a conversation — with its uploaded files and a portable context of your own — between ChatGPT, Claude and Gemini, entirely locally: no backend, no account, no telemetry. Six layers whose import boundaries are enforced by the linter and CI rather than by review; a typed message contract that fails to compile on a missing handler; DOM selectors shipped as remote config, so a platform restyle is fixed without waiting out a store review | TypeScript · Chrome Extensions · Playwright |
| car-body-type-classifier 🤝 🔗 | Classifies vehicles into 8 body types — EfficientNetB3, 0.94 test Macro F1 on 12,250 curated images under a strict 95MB model-size constraint. I owned the data side: async scraping pipelines, multi-source dataset assembly and deduplication, and the EfficientNetB0/V2S baselines | Python · TensorFlow · Playwright |
| geo-news-scraper 🤝 🔗 | Tracks local news across Kocaeli: Scrapy spiders over 5 sites, rule-based NLP classification, deduplication by sentence-embedding similarity, and a 4-layer geocoding pipeline (regex → gazetteer → spaCy NER → Google Geocoding). I built the React + Google Maps dashboard with live filtering | Python · FastAPI · MongoDB |
| product-genai-pipeline | Enriches bare e-commerce listings: description generation (T5), category prediction (TF-IDF + logistic regression), clustering of similar products (KMeans over TF-IDF and sentence embeddings), and long-text summarisation, with the full preprocessing and export pipeline around it | Python · Transformers |
🤝 = built with teammates · 🔗 = repo lives on a teammate's account
More projects — coursework and team work outside the AI/NLP line
| Project | What it does | Built with |
|---|---|---|
| wordCrush-mobile 🤝 | An iOS word puzzle game with 8-directional swipe input, combo scoring and chain-reaction power tiles, backed by a Trie over 59K Turkish words for O(m) lookup and a background Isolate solvability solver | Flutter · Dart · Riverpod · ObjectBox |
| CargoOperationsSystem 🤝 | Plans delivery routes over a real road network, solving a vehicle routing problem with the Clarke-Wright savings algorithm | FastAPI · React · Supabase |
| GraphCite 🤝 | Explores citation networks between papers — H-Index, H-Core, H-Median — through an interactive D3.js graph in a Swing/JavaFX hybrid UI, with a custom JSON parser | Java · JavaFX · D3.js |
| lidar-geo-analysis 🤝 | Picks walls and corners out of 2D LiDAR point clouds with RANSAC, converts polar to Cartesian, and exports an SVG map. Covered by CTest unit tests | C++ · CMake |
| student-schedule-system 🤝 | Exam scheduling, seat planning and classroom/student management for a university department | PyQt5 · PostgreSQL |
| diabetesTrackingSystem | Desktop app for patients and doctors to track blood sugar, diet and exercise, with alerts and a doctor-patient dashboard | PyQt5 · SQLite · PostgreSQL |
| vehicle-security-and-control-system | Arduino simulation of a car's safety electronics — seatbelt interlock, automatic headlights, temperature and fuel monitoring on an LCD | Arduino · C++ · Proteus |
| sosyal-kutuphane-platformu 🤝 🔗 | A social cataloguing app for books and films — follow system, activity feed, the works | React · Supabase |
| TowerDefenseGame 🤝 🔗 | A 2D cyberpunk tower defense game — tactical squads with distinct combat mechanics | Unity · C# |
Languages
ML & NLP
Frameworks & Libraries
Platforms, Databases & Tools
Commencis Cloud Camp (AWS) · Üretken Akademi Kılavuz Kariyer (NVIDIA — AI Fundamentals & GPU Architecture) · Boğaziçi University DataCamp (NLP / CV / Big Data) · GSU Digital Leadership Academy · Schneider Electric Global Student Experience