AI engineer in the making, and a senior Artificial Intelligence & Data Engineering student at Istanbul Technical University (İTÜ). I train language models from scratch, fine-tune vision foundation models, and make robots pick things up in simulation.
🔎 Open to full-time & part-time roles in ML / AI Engineering, Data Science, Computer Vision, Robotics and ML Research. Graduating January 2027.
Does HaLViT's "half of the weights are enough" W+Wᵀ trick (CVPR 2024 W) carry over from Vision Transformers to GPT-style language models? Advised by Prof. B. U. Töreyin. I've led the project since September 2026.
- Built a controlled experiment ladder: GPT-style LMs trained from scratch on WikiText-103 (614M tokens per run). Only the sharing mechanism changes between arms, so every perplexity delta can be attributed to it.
- Quantified the "sharing tax": attention-only W+Wᵀ cuts stored weights by 16.7% for a +3.97% perplexity cost, the cheapest trade-off of all variants (3 seeds).
- Showed that, at fixed storage, looping layers beats W+Wᵀ, both on language modelling and on synthetic multi-hop reasoning.
- Ran it like real research: pre-registered hypotheses, multi-seed runs, VRAM pre-flight checks, and checkpoints versioned on Hugging Face. 11 merged PRs.
PyTorch · Transformers · bf16 training · experiment design
| Project | What it does | Stack |
|---|---|---|
| 🗄️ tiny-sql-rl | Solo project, follow-up to Voice-to-SQL. Trained Qwen3-1.7B with GRPO on an execution reward (4-bit QLoRA, one 8 GB laptop GPU) to replace SQLCoder-7B. Spider dev execution accuracy 54.1% → 59.0% (official script, 84 wins / 33 losses). On spoken questions it ties the 7B model from clean audio to 10 dB SNR. Re-measuring the old pipeline showed its 14 s "LLM bottleneck" was a runtime problem: SQLCoder takes ~2 s when run properly. Released on Hugging Face with a GGUF build and a local voice demo. | TRL, Unsloth, vLLM, llama.cpp, Gradio |
| 🎙️ Voice-to-SQL | Solo project. Speak a question about a database and hear the answer: Whisper ASR → SQLCoder-7B (4-bit) text-to-SQL → validated SQLite execution → spoken reply via VITS/gTTS. Benchmarked on 400 noisy Spider queries (clean to 0 dB SNR): WER rises 10× at 0 dB but SQL execution accuracy drops only ~7 pp, and profiling shows the LLM accounts for ~96% of latency. (A follow-up, tiny-sql-rl, traced most of that to the runtime setup rather than the 7B model.) | Whisper, Transformers, SQLite, Coqui TTS |
| 🎯 SAMet: Video Object Segmentation with SAM 2 | Team project, İTÜ Computer Vision course. Evaluation and fine-tuning pipeline for Meta's SAM 2 on DAVIS 2017: zero-shot baseline (0.905 J&F), mask-decoder fine-tuning with a frozen backbone, and ablations over LR scheduling, boundary loss and uncertainty-weighted losses. | PyTorch, SAM 2, DAVIS, TensorBoard |
| 🧠 Multi-Modal Brain Tumor Segmentation | Team project. Comparative study of U-Net, LSTM-UNet and Transformer-based models for segmenting tumour regions in multi-modal MRI (T1, T2, FLAIR, T1Gd) from the BraTS 2021 dataset. | PyTorch, Jupyter, medical imaging |
| 📈 Unsupervised Anomaly Detection in HDFS Logs | Pair project with Yusuf Karamustafa (pair-programmed; commits are under his account). End-to-end pipeline: parses 11M+ log entries (1.47 GB) with Apache Spark, maps raw messages to 29 event templates, builds 575K block-level traces, and trains an LSTM autoencoder that flags anomalous executions without labels. Includes threshold sensitivity and scalability benchmarks. | Apache Spark, PyTorch, LSTM |
| 🤖 Robotics Med Cup Collector | Team project. Autonomous pipeline for a Hello Robot Stretch in MuJoCo: navigate to tables, detect humans and cups, grasp with visual servoing, and drop the cup in the trash. | MuJoCo, YOLO, Python |
- 🗣️ Turkish–Italian Idiom Detection: team project. Multilingual sequence-labelling model that tags idioms in text, combining a pre-trained transformer (BERT / RoBERTa / XLM-R) with dilated convolutions, BiLSTM, multi-head attention and a CRF layer.
- 📐 Point Set Registration & Image Stitching: SVD (via power iteration) and Kabsch–Umeyama written from scratch in NumPy, used to align 3D point clouds and stitch 2D images with rotation, translation and scale.
- 🎬 TMDB Movie Analytics Dashboard: team project. Data cleaning and EDA in Python, feeding an interactive Power BI dashboard on budget vs. revenue, production geography and actor co-star networks.
Languages: Python · TypeScript/JavaScript
ML / DL: PyTorch · Hugging Face Transformers · TRL (GRPO) · Unsloth · vLLM · llama.cpp · Ultralytics YOLO · SAM 2 · U-Net · Whisper
Data: Apache Spark · NumPy · Pandas · Jupyter · Power BI
Tools: Git · uv · CUDA · TensorBoard · MuJoCo

