Production Stack
Core Architecture
Parses code using tree-sitter into functions, classes, and methods — not arbitrary character splits. Preserves semantic boundaries.
Maps function call graphs so when Function A calls Function B across files, the system knows they're related — even in different repos.
Combines keyword precision (BM25) with semantic understanding (Gemini 1536-dim embeddings) using Reciprocal Rank Fusion.
Caches LLM responses by query similarity (cosine ≥ 0.92). Cuts API costs by ~70% and reduces latency from 3s → 100ms.
A stateful multi-step agent that retrieves, grades context quality, and reruns with a better query if the first answer is poor.
Automated benchmarking for Faithfulness, Context Precision, and Answer Relevance. Real numbers on your resume.