Applied Machine Learning

The argument

The hardest call in applied ML isn't picking a model. It's deciding whether to build, buy, or fine-tune in the first place, and that decision lives at the intersection of cost, latency, data contracts, and what your team can actually maintain six months from now.

The pillar I work in answers that question with working systems. A reference page that ships six computer vision capabilities client-side and tells you what each costs in browser, cloud, and fine-tune tiers. A logistics ML lab that exposes a stable prediction API designed to swap from a lane-average baseline to a Random Forest without breaking the contract. A capital markets sandbox with a working price-time-priority matching engine that bots can actually trade through.

The thesis under all of it is the same: build/buy/fine-tune is a procurement decision, not a research decision. The people who can frame it that way and back the framing with working systems are the ones who get to make those calls at scale.

I reason about ML systems end-to-end. From cost-tier selection and data contracts to matching engines and model swap-in paths.