I gravitate toward operations that break at scale.
Not because they are dramatic. Because the failure is almost always structural, and structure is fixable.
Why product and operations
I did not arrive at product from design or engineering. I arrived from operations — transitions, escalations, SLAs, the parts of a business that only become visible when they fail. That background changes what I notice. When a team tells me annotation quality is slipping, my first question is not about the annotators.
Product without operations produces roadmaps that assume the machine runs itself. Operations without product produces heroic firefighting that never becomes a system. The interesting work is in between, and there are not many people standing there.
Enterprise transformation taught me the part that transfers: at sufficient scale, anything that depends on a person checking it has already failed. You just have not found out yet. That lesson moved cleanly from telecoms platforms to perception data — the volumes changed, the failure mode did not.
What I believe about production AI
Data is the product surface nobody owns
Models get product managers. Datasets get a spreadsheet and a vendor contact. That asymmetry is where most production AI quietly degrades.
Quality is an architecture, not an inspection
You cannot inspect quality into a dataset at a million frames. It has to be specified upstream and enforced by the workflow.
Cost per frame hides the real bill
The expensive frame is the one that comes back twice. Rework is set by how clear the specification was, not by the hourly rate.
Human-in-the-loop is an operating model
It is not a fallback for when the model is unsure. It is a system with routing, economics and a throughput ceiling of its own.
What I am building toward
Autonomous systems are the hardest version of the data problem — the physical world does not agree to a taxonomy, and the edge cases are the entire point. Everything I am interested in next sits there: simulation feeding annotation, active learning deciding what gets collected, and the loop between production failure and the next dataset closing without a human in the middle of it.
What I do not do
I am not a foundation-model researcher or a perception-algorithm engineer. My work sits at the layer where data, product, workflow, vendors and quality meet — which is usually where production AI actually breaks.