Updated July 2026
Same problem as always: take a person’s intent, a query, a click, a half-formed preference, and resolve it to the right next step. Right now most of my attention is on the two parts of that work most teams skip.
The first is the catalog. At Velou I spend real time turning thin, messy ecommerce product data into one real ontology, item by item, because a shopper’s intent can only reach a product that’s described well enough to be found.
The second is evals: golden datasets, regression suites, offline metrics that survive contact with production, LLM-as-judge setups I’ve calibrated rather than trusted on faith. The evals are what let the system learn instead of merely staying live. I’m putting models to work where they earn their place, and the eval layer keeps that honest. I’m careful about the line between a convincing demo and a system I’d put in front of a customer.
What I keep circling is where intent actually lives. In commerce it surfaces as a query or a click, thin and late. The richer signal sits earlier, in what someone does before they ever type. How much of that you can read, and how reliably, is still an open question.
Reading lately: evaluation harnesses and postmortems. The bottleneck on putting a model anywhere load-bearing is trust, and that is a measurement problem before it’s a model problem.