Built to find the failure mode,
not just the demo.

Naive RAG misses structurally adjacent evidence. FAISS ANN search with NetworkX two-hop re-scoring, tested across three pipeline variants. Outperformed the naive baseline on all four RAGAS metrics.

Fixed vocabulary-mismatch failures in pure semantic search with hybrid sparse-dense retrieval, raising recall@10 35% over baseline and reducing latency from sub-600ms to sub-100ms.

Evolved the compiler beyond model-specific prompt tuning into a general context layer, transferring a user's own knowledge and project information into agent-ready structure for agentic workflows.

Deployed an agentic interview system of three reasoning agents, transcribing spoken answers via Whisper to shape follow-up questions in a candidate's real experience. University of Hertfordshire accredited.