Bekir Sadik
Altunkaya

Leading distributed teams on reinforcement learning from human feedback and post-training pipelines for large language models.

Led training and fine-tuning efforts for advanced language models. Oversaw code quality across 10+ projects spanning RLHF, SFT, and rubric data pipelines generating over $10M in cumulative revenue.

Reviewed and approved 600+ large-scale GitHub PRs for code generation models. Evaluated complex code for reasoning depth, syntax correctness, and multi-file context handling.

Engineered an end-to-end hyperspectral imaging pipeline processing 9.2GB+ HDF5 data cubes. Built a spectral unmixing engine achieving 97-99% R² fit. Optimized a YOLO-based CV system on a 15k dataset.

Shipped a fast video player using YouTube infra and Next.js across web and mobile study apps. Reached 100K+ monthly learners and drove 2M+ video plays/year. Built a supplementary vanilla-JS version so schools and older devices could run the same player without a framework.

Engineered secure Mobile Signature solutions compliant with XAdES and XML standards. Optimized XML signature parsing, reducing validation latency by 35% to support 2.5M+ daily transactions.

TÜBİTAK 3005 project. Built a 4-stage multi-modal semantic relevance detection pipeline using a dual-agent Vision-Language Model architecture (LLaVA v1.6) with sliding window context. Achieved 83.3% precision and 0.79 F1-Score. Published at SIU 2026.