Project · Flagship
Agentic Analytics
An LLM agent that runs a real analytical workflow over public data: it plans, queries, and critiques its own output with an LLM-judge before it answers — returning a cited, verified narrative.
The problem
Most "agentic analytics" demos answer one question, once. The version that matters runs a real workflow and can tell when it's wrong. Wiring an agent to a database is the easy part; the work is verification.
The approach
Plan → act → verify, over public data (NCES/IPEDS). The interesting design surface is the eval: decomposing accuracy into retrieval ceiling × judge conversion, then engineering each independently.
Verification, front and center
An LLM-judge grounded in domain logic, adversarial refutation, and grounded citations so a human can check the work. Full method in the writeup: Decomposing agent accuracy.
Try it
Coming soon: a 2–3 minute recorded demo, the source repository, and a gated live demo for those who want to run it hands-on.