My Work

Voice-first clinical training simulator — I'm co-founder and CTO. Medical students play the doctor: patients arrive, you talk to them in real time, order tests, treat, disposition. An attending physician agent watches every decision and grades it against the published guideline (NICE, ESC, AHA, GINA, GOLD). Live at medkit.clinic, invite-only.

Source-grounded Quran Q&A with hybrid RAG. 6,236 verses, 15 translations and 6,237 tafsir chunks in a pgvector store with HNSW indexing; tsvector lexical search (Turkish diacritic-insensitive FTS) fused with semantic search via Reciprocal Rank Fusion. The LLM never produces scripture — Arabic verses are always served from the data layer with references, so it can't hallucinate. Live on the App Store.

Your own computer's coding agent, from your phone — not a hosted copy of it. A daemon on your machine drives the real claude or codex process in your repo, signed into your own account, and speaks WebSocket over your tailnet. Approvals land on the lock screen with the exact command; turns keep running when the phone loses signal. Ships with a desktop panel that watches every machine at once.

A study tool for the Dutch motorcycle theory exam (CBR category A), in Dutch, Turkish and English. Runs as a web page with no build step and as a React Native app. Vocabulary cards, road-sign drills, and a mock exam that follows the real format — one block, 50 questions, 30 minutes, 41 to pass.

A cannulation trainer with a C++ core in Unreal Engine 5.8: procedural skin, arm and hand geometry, patient profiles with different vein anatomy, and a guided demo mode. A WebGL build shares the same rules engine, so the browser and the desktop grade identically.

An electric care bed with a proprietary remote and an app that barely worked. I reverse-engineered the Richmat WiLinke W1 protocol, wrote it up, and built an iPhone app plus an ESP32 BLE-to-WiFi bridge so the bed can also be driven from a PC — or by anything else that can send an HTTP request.

Real-time automated trading bot consuming Binance WebSocket streams and executing custom strategies. Deployed for banking clients (DenizBank, Akbank). My code trades better than my instincts.

End-to-end YOLOv5 detection pipeline for ship electric charging sockets with auto-locking. Anchor optimization, real-time inference. Published in IEEE.

AI-powered embryo quality pre-labeling from time-lapse imagery for physician decision support. CNN classifiers (ResNet-50, Inception-v3, DenseNet-121, EfficientNet) with transfer learning on clinical data — deployed in collaboration with Acıbadem fertility clinics.

Master's thesis (defense stage). Multi-modal deception classification fusing gaze patterns (fixation/saccade) with temporal facial micro-expression features. Sole researcher.

Computer vision and ML for hospital decision support — illness detection with YOLO and CNN classifiers, deployed in production with a 10-person agile team and clinicians.

Symptom-based disease prediction system with an optimized model selection pipeline. Cross-validated classifier benchmarks (Decision Tree, Random Forest, SVM).

Mobile app that recommends playlists based on your current mood, detected from selfies. Trained an emotion prediction model. The AI says I always look tired.

State notification system for Burger King, Watsons, Starbucks — automating official employee and business notifications. AI-driven government grant calculations. Microservices with fault tolerance.