Studying at Codam College (42 Network) — based in Amsterdam, NL.
Lexical-first retrieval (BM25) fused with a schema-constrained knowledge graph — Qwen3-0.6B + Outlines extracts entities/relations into Neo4j, so graph traversal pulls in related chunks BM25 never lexically matched.
Retrieval-augmented generation over the vLLM codebase — custom identifier-aware tokenizer, AST-based chunking, tuned BM25 (k1/b grid search), IoU-scored recall@k eval, grounded answers via Qwen3-0.6B.
Function calling for LLMs via constrained decoding — masks invalid tokens at the logit level so a 0.6B model emits 100% schema-compliant JSON.
Interactive lab for learning math, algorithms, data structures and concurrency visually — live curve transforms, step-through algorithm playback, and a deterministic thread scheduler for reproducing real race conditions and deadlocks.
Dining-philosophers variant in C: threads compete for shared resources under FIFO or earliest-deadline-first scheduling, with burnout deadlines.
Let's talk about your project and how I can help you.