Other

Three of the systems I have built had the same constraint from three different directions: the machine doing the inference could not rely on a network to somewhere else. It is the least fashionable constraint in the industry and it changes almost every decision downstream of it.

McDonald's: models at the edge of a restaurant fleet. Each restaurant ran its own local model rather than calling a central endpoint, which is the only shape that survives the connection being the least reliable component in the system. I designed the data gathering and training path around it: a distributed MiNiFi deployment that projected out PII and unwanted fields before anything left the building, compressed and batched what remained, and queued whatever it could not send until it could. Each site exposed its model endpoint and collected the classifications it made, which were evaluated automatically, so the training loop closed without anyone having to visit a restaurant.

A government intelligence agency: training inside an air gap. A T5 translation model, trained in an environment with no route out. The model is the least interesting part. The constraint is the whole job: every dependency, base checkpoint and tool has to be carried in deliberately, and there is no second trip for the one you forgot.

Metro: two servers in an airport DMZ. ~400 cameras, sub-second, air-gapped, on hardware that was never going to grow. Written up in full on its own page. Same shape as the agency work, from the other side of the table.

The infrastructure underneath. All three major clouds (GCP, AWS, Azure), with and without Terraform. Kubernetes as the default deployment target, Docker, BigQuery. Plus a working sense of what actually changes when the target is air-gapped or customer-controlled, which is rarely the architecture and almost always everything around it: how dependencies arrive, how you ship an update, and how you find out something broke.

Created the org's internal learning path for Googlers who wanted to learn AI from first principles. Grew to ~500 members worldwide. Awarded Initiatives Champion by the org for it. Also conducted ~100 technical interviews for L4-L5 ML and software roles.

Founded one of the first university AI labs in Argentina at a time when there was little local material: the lab ran as a paper-reading group where members studied and taught each other current work. The research project was with a hospital that lent us camera recordings: a generic system for real-time multi-camera tracking of moving objects mapped into an abstract 3D representation, with a batch layer for querying journeys and flows after the fact. Think product analytics, but over physical space and the people moving through it; for the hospital, that meant finding inefficient paths. Metro Smart Cities spun out of this lab, and that research became a shipped product: Metro sold it as metro follow, described on its own site as Google Analytics for physical retail.