Work
Client projects our founder led before starting Keelwork. They are under NDA, so the clients are not named and identifying details are left out. The story, the architecture and the measured results are real.
RAG on your documents
Answering questions about 10,000 contracts without making them up
Contract question-answering that cites the clause it used and says 'I don't know' when the contracts don't cover it.
- 92%answers correct and cited
- 1.3%answers with an unsupported claim
- $0.012cost per answer
AI agents
A support assistant that knows when to say "I don't know"
A support copilot that drafts grounded, cited replies from the knowledge base and resolved tickets, auto-sends what it is sure of, suggests a draft to an agent when it is useful but unsure, and escalates what it should not answer.
- 42%conversations resolved without a human
- 0.6%replies with an unsupported claim
- 4.4/5customer rating on AI-handled replies
Evals, cost and latency
The AI feature worked in the demo and broke in production
A shipped AI feature that worked in the demo and failed in production.
- 92%answer accuracy, now measured
- 2.3sp95 latency, from 4.1s
- 35%lower cost per answer
AI in existing systems
Adding a copilot to software that was never built for one
Adding an AI assistant to an existing enterprise system without touching the system of record.
- 0records surfaced beyond a user's access
- 100%retrievals filtered by the caller's identity
- 0customer data leaving the region