I keep going one layer down.

The order matters. Building products taught me that most hard problems are really data-model problems wearing a user-interface costume. Building AI products taught me that the interesting engineering is in routing, permissions, and metering, not in the model call. Writing a neural network framework from the autograd up taught me what that model call is actually doing.

And writing a kernel taught me that I had been describing several layers of my own stack from memory rather than from understanding. That is the honest reason ChronoOS exists: not because the world needs another operating system, but because I did not like how vague my answers got below the runtime.

The pattern in the work is not the technology. It is that every project is one abstraction level below the thing that made me curious. That is also why I contribute to open source: reading a large codebase you did not design, finding the actual cause of a bug, and getting a change accepted is a different skill from building something from scratch, and it is the one that transfers.

I work part-time as a software engineer at Helix Grid, building AI and LLM features. Most of that work sits around the model call rather than in it: context retrieval, output validation, fallbacks, and an evaluation suite.

I also have one research study in progress, on whether training-time masking makes tabular classifiers hold up when the data arrives with holes in it. Nothing is published, and the page says exactly that rather than implying a result I do not have.