Role
Founding engineer — RAG pipeline, evaluation framework, prompt system, and frontend.
Published
2025 — present
- Python
- FastAPI
- pgvector
- Claude API
- Next.js
Impact
- -52% first-response time
- 94% answer acceptance rate
- 40k drafts generated monthly
Most AI copilots fail the same way: they hallucinate confidently and erode the team’s trust within a month. Lumen was built around the opposite bet — that a slower, citable, verifiable answer beats a fast, fluent guess.
Architecture
The core is a retrieval-augmented pipeline tuned for support conversations, not general questions:
- Grounded generation. Every draft cites the exact doc sections it used. Agents see sources inline; if the retriever finds nothing, the system says so instead of improvising. “I don’t know” is a feature.
- Evaluation harness. I built an offline eval set of 600 real past tickets with gold answers. Every prompt or model change is scored on factual accuracy, tone and citation quality before shipping — no vibe-based releases.
- Tone transfer. The system mirrors each agent’s own past replies as a style reference, so drafts sound like the team, not like a chatbot.
- Human in the loop. Nothing sends automatically. The agent edits, approves or regenerates. Usage logs showed agents keep Lumen’s wording ~94% of the time while still editing nearly every draft — exactly the collaboration mode we designed for.
Results
First-response time dropped 52% across a 200-agent support organization. More importantly, new agents reach full productivity in weeks instead of months, because the system is a living training corpus, not just an autocomplete.
What I’d do differently
Ship the evaluation harness earlier. The first month of prompt iteration without it was guesswork dressed up as progress.