Robin, market research
A market-research agent in production. It answers only from retrieved data, and an AI judge reviews every answer.

The context
Robin lives inside a financial analysis product. The idea: let users ask, in plain language, about the same data they already see on the platform.
A generic chatbot wouldn't do: in finance a made-up answer is worse than none, and the agent must never give investment advice.
What I built
An agent that plans, runs tools over SQL views and search, and replans when needed. It asks for clarification when a question is ambiguous.
Online evals: a sample of real answers goes through an AI judge that checks each answer is grounded in what was retrieved.
How
Plan and execute
Router, planner, ReAct executors and a replanner in LangGraph. Every step is traced.
Retrieved data only
The agent answers with what it finds or says it doesn't know. No investment advice.
Measure in production
Groundedness judge, tool-error detector, 120+ reference cases from human review and regression tests in CI.
Got a similar process?
Tell me about it. The first call is free.