Me

I work as an AI and ML engineer. Most of my time goes into the parts of a system that have to survive contact with real traffic: training embedding models, serving LLMs, building retrieval over large collections, and the cloud plumbing that keeps those pieces talking to each other.

I also lead and mentor small teams through that same path. A sketch is cheap. Getting something into production — with latency you can quote, failure domains you can isolate, and a cost that does not explode — is the part I care about.

I am interested in retrieval that actually grounds an answer, not a fluent guess. That means embeddings that fit a domain rather than a generic average, graphs over documents when a flat search is not enough, and feedback loops so the system gets less wrong over time. The same curiosity applies when the input is not text: time series, faces, gaze, audio, video. Mixing those signals is more interesting to me than treating language as the only modality.

I started in biotechnology and mathematics and spent the last decade on applied machine learning. The thread is the same: noisy measurements, a model that has to be honest about uncertainty, and a system around it that other people can run.

On the side I build tools that watch video — live streams and clips — and flag events as they happen. It is a playground for detection, tracking, and small vision-language models, without the ceremony of a client project.

I like problems where the interesting work is in the middle: not a demo, not a research paper, but the distance between the two.