Selected Work
2025 · AI
Atlas AI
A retrieval-augmented assistant that reads 500k enterprise documents and answers in under 800ms — grounded, cited, and auditable.
Atlas AI
Client
Atlas Labs
Role
Founding Engineer
Year
2025
Discipline
AI
(01)
Problem
Atlas had a working LLM prototype that impressed demos but failed in production: hallucinated citations, 12-second response times, and zero observability into why answers were wrong.
(02)
Solution
Built a hybrid retrieval stack (BM25 + pgvector) with citation-required prompts, an eval harness that runs 200 test cases on every deploy, and a trace explorer so every answer links back to source paragraphs. Single GPU node, sub-800ms p95.
(03)
Outcome
Adopted by 4 Fortune-500 customers in 6 months. Hallucination rate dropped from 18% to 0.4%. Average response time cut by 93%.
800ms
P95 latency
0.4%
Hallucination rate
500k
Documents indexed
4
F500 customers
Stack
PythonpgvectorLangChainFastAPIVercel AI
Next case