About

I like the unglamorous half of machine learning: the serving path, the retry, the cache, the benchmark nobody wants to write.

I started in research — object detection, dark-web threat modelling, campus parking forecasts — and learned quickly that the model was rarely the hard part. Getting predictions out of a notebook and into something a business can rely on, at a latency and cost that survive scrutiny, is where I have spent the eight years since.

At BILL that has meant agent frameworks over financial documents, a grounding standard the whole org benchmarks against, and an OCR stack whose tail latency I took personally. At home it means an inference server I am writing by hand, a photo NAS that replaced a subscription, and a tracker for a baby who has not arrived yet.

I write things down as I go, mostly to catch myself claiming a speedup I never measured. Those notes live under writing.