An answer is only useful when you can trace where it came from.
Enterprise RAG platform for internal company procedures (1k–10k documents per deployment) — single-tenant, on-premise or customer-cloud, with an optional fully self-hosted LLM so no document ever leaves customer infrastructure.
Hybrid retrieval pipeline: pgvector cosine search + PostgreSQL full-text search fused via RRF, then cross-encoder reranking — p95 retrieval latency ≤ 700 ms before the LLM call. Every answer cites source document and page; insufficient context returns an explicit refusal instead of a hallucination.
Security-first architecture: department isolation enforced by Postgres Row-Level Security (not application code), ACL filtering before any content reaches the model, and an append-only audit log (no UPDATE/DELETE path) designed for EU AI Act Art. 12 logging requirements, plus prompt-injection hardening.
Built a deterministic synthetic Polish eval corpus (107 documents, 94 golden test cases including injection canaries) for regression testing of retrieval quality.
- p95 retrieval ≤ 700 ms
- Postgres RLS + ACL
- EU AI Act Art. 12 audit log
- 94 golden test cases