01 · The challenge
What needed solving
Underwriters spent 40+ minutes per quote hunting through 12,000 policy PDFs, internal memos and case-law summaries — costing throughput and producing inconsistent decisions.
02 · Our approach
How we framed the work
We started with a discovery sprint to map the user journey, business goals and real constraints. From there we wrote a fixed-scope plan: clear milestones, weekly review gates on a staging URL, and a written exit criterion for every phase. The ai/ml space rewards teams that ship — not teams that plan — so we biased the engagement towards working software from week two onward.
03 · The solution
What we built
Built a hybrid-retrieval RAG pipeline (pgvector + BM25), Claude 4.6 Sonnet for reasoning, a LangGraph agent that asks clarifying questions, strict citation-only answers, and an internal eval harness with 240 golden questions. Live cost + latency dashboard for ops.
04 · The results
What changed for the client.
Underwriter research time dropped from 40 minutes to 11 minutes per quote
94% answer accuracy on the golden eval set, every answer with source citations
Cost per query held under $0.08 via prompt caching and model cascading
Tech stack
Want a similar outcome for your business? We'd love to hear about it.