İsmet İnan

A GPT-style transformer decoder built from scratch, to understand the internals rather than call an API.

• Multi-head self-attention with causal masking and learned positional embeddings
• Temperature, top-K, and nucleus (top-P) sampling for controlled generation
• Custom tokenizer with a JSON vocabulary, kept separate from the model code

A deployed platform where Bilkent students share course materials and discuss courses in threaded conversations.

• Go API with JWT auth restricted to verified Bilkent students
• Postgres 15 behind Goose migrations, containerised with Docker Compose
• Media uploads to AWS S3, DNS on Route 53, transactional email via Resend

Cohere Labs — Tiny Aya Expedition · team project

A multilingual routing layer that sends each prompt to the local model most likely to answer it well, instead of paying to run every model.

• Four-gate pipeline: FastText language ID, a speculative A/B probe when confidence is low, reward and safety scoring, then a hard fallback model
• Routes across GGUF models served by llama-server, with coverage-based language-to-model mapping
• Parses token logprobs into confidence and margin metrics to decide when probing is worth the extra inference

An AI mentorship app for self-taught artists, giving structured feedback on drawings where there would otherwise be no mentor at all.

• FastAPI and PostgreSQL backend with a swappable AI provider layer, so the model behind the feedback can change without touching the app
• Flutter client built around Mentors, Lessons, and Profile
• Redline consistency tests to check that the critique a drawing gets is stable rather than arbitrary