Computer Engineering student at İstanbul Medeniyet University (class of 2027) · Machine Learning · LLMs & Agents · Time Series
- 🥇 1st place (individual), Türk Telekom AI Camp 2026 — selected among 30 finalists from 4,464 applicants. Won with BundleKom, a telecom customer-analytics and RAG / tool-calling copilot built on 200k customers × 36 months of data.
- 🏅 Top 5, Trendyol–TEKNOFEST E-Commerce Hackathon 2026 — with team DualCore: 5th of 377 teams in the Kaggle stage and 5th among the 10 finalists. Search-relevance system with two Gemma-4-12B LoRA adapters fused by a HistGradientBoosting decision model.
- LLMs & agents — RAG, tool calling, LangGraph agents, LoRA fine-tuning (Llama 3.1, Gemma), knowledge-graph-grounded decision support.
- Time series & predictive maintenance — RUL prediction, industrial anomaly detection (Anomalib, MVTec-AD), explainable AI, representation probing.
- Turkish NLP — BERTurk / XLM-R classification, safe domain assistants, search relevance and similarity analysis.
- End-to-end delivery — FastAPI + PostgreSQL + React/TypeScript backends, MLflow experiment tracking, Docker.
| Where | Role | When | What I built |
|---|---|---|---|
| Ümraniye District Governorship | Intern | Aug – Sep 2026 | Media-monitoring system (FastAPI, PostgreSQL, React/TS) with duplicate detection, role-based access and explainable Turkish sentiment / risk / category analysis with alerts and daily reports |
| ROKETSAN STAGE | Software & Digital Design Intern | Jul – Aug 2026 | Python assembler that parses DMA instructions from hardware documentation and emits binary / hex machine code, with syntax and rule validation and unit tests |
| BekTech | Machine Learning & LLM Intern | Mar – May 2026 | Predictive maintenance and RUL on multivariate time series; benchmarked ML, deep learning and foundation models by RMSE and NASA score; LLM-based decision-support prototypes |
| Project | What it does | Stack |
|---|---|---|
| Cüzdan Koçu — AI Family Finance Coach | Family budgeting with goals and envelopes, receipt OCR and role-based data access; a LangGraph agent with tool calling, user-approved actions, SSE streaming and speech-to-text / text-to-speech. | Next.js · FastAPI · PostgreSQL · MinIO · LangGraph |
| Sina — Turkish Health Assistant | Routes patient questions to one of 16 medical specialties on 48,816 balanced samples (BERTurk, 0.69 macro-F1) and answers with a Llama 3.1 8B model fine-tuned via 4-bit LoRA, with PII cleaning and safety filters. Model on Hugging Face. | PyTorch · Transformers · Unsloth · Ollama · Streamlit |
| CopyGuard — Turkish Similarity & Plagiarism Analysis | Hybrid similarity over PDF / DOCX / PPTX / TXT / Markdown combining MinHash, TF-IDF, embeddings and block overlap; live progress over SSE and CSV / HTML / SVG / PNG reports. | FastAPI · Next.js · SQLAlchemy |
| MovieMind — Hybrid Movie Recommender | Content-based, item-based collaborative filtering and Apriori association rules behind a full-stack app with profiles, friends and TMDB integration. | React/TS · Node/Express · Flask · scikit-learn |
| Appointment No-Show Prediction | LightGBM + SMOTETomek + threshold tuning on 110k medical appointments, 83% recall on no-shows. Live demo. | scikit-learn · LightGBM · Streamlit |
| ŞehirPulse AI | Citizen complaint intake that classifies requests and routes them to the right municipal unit; CI, Docker and a Scrum team workflow. | FastAPI · React · SQLite · Docker |
| Emergency Drone Coordination | Multithreaded C client-server simulator: thread-safe lists, free-list memory management, TCP/JSON protocol, heartbeat and an SDL2 real-time viewer. | C · pthreads · SDL2 · json-c |
| Multi-User Communicating Shells | GTK4 terminal where parallel shells exchange messages over POSIX shared memory; pipes, redirection and history. | C · GTK4 · POSIX |
- BundleKom — telecom customer lifecycle, usage and complaint analytics on DuckDB; churn model validated over time with PR-AUC, Precision@K, Lift@K and calibration; RAG + tool-calling copilot with ChromaDB. Türk Telekom AI Camp 2026, 1st place.
- Trendyol–TEKNOFEST search relevance (DualCore) — hard-negative mining supervised by an LLM, two Gemma-4-12B LoRA adapters with different error profiles, HistGradientBoosting fusion with explainability. Top 5 of 377 teams.
- Turbofan RUL with foundation models — zero-shot Chronos / TimesFM, frozen-backbone heads, LoRA and full fine-tuning vs. LSTM / BiLSTM / CNN-1D / Transformer on C-MAPSS FD001–FD004 and PHM08; representation analysis with linear probing, PCA / t-SNE and latent-direction stress tests.
- LLM + knowledge-graph fault diagnosis — FAISS RAG, NetworkX symptom–cause–solution graph and a multi-agent flow on 4-bit Mistral-7B-Instruct with retrieval verification and source-controlled answers.
ML / AI
Backend / Data
Frontend / Systems
- Kairu — Community Relations Manager (Aug 2025 – Mar 2026). Coordinated AI / data-science training partnerships with university clubs; taught a hands-on MLflow experiment-tracking and model-versioning session at AIdea Project Camp.
- SistersLab — AI Bootcamp Assistant, volunteer (May – Jul 2025). Reviewed participant assignments and mentored on machine learning, Python and NLP.