Projects
I keep one foot in the build so my AI strategy isn't guesswork. Here's a working system I built to pressure-test where agentic analytics actually breaks — the same question every exec should ask an AI vendor: how do you know the answer is right?
An LLM agent that answers questions over a real e-commerce warehouse and won't ship a number it can't ground in the data: it plans, runs read-only SQL, and passes its answer to an independent judge before replying. Backed by a 26-question eval — including traps a naive query gets wrong — that measures accuracy and groundedness separately. Wiring an agent to a database is the easy part; the work is knowing when the answer is wrong.
A fan's look past the punditry, run like a real analysis: 25+ questions about Kevin O'Connell's offense, each with a hypothesis written before the data, a league benchmark, and honest uncertainty. Also an experiment in how much of data science to hand an AI agent.
Additional experiments in agentic analytics and LLM evaluation will land here.