The model proposes. My code decides.
A language model reads a customer message and proposes an order. A validation gate — ordinary, testable code — decides whether it may exist. Edit the parsed order and watch the gate refuse it.
I write SQL against real schemas.
Multi-tenant, ownership-aware, and enforced in the database as well as the API. Write and run your own query against 1,002 rows, on a SQL engine I wrote by hand.
Production systems, not demos.
Leases, retries, dead-lettering, heartbeats. The interesting part is never the happy path — kill a worker mid-lease and trace how the job comes back.
Retrieval, not the buzzword.
Similarity is not relevance. Past a certain score you retrieve noise that reads plausibly, which is worse than retrieving nothing. Move the cutoff and watch boilerplate cross the line.
I turn data into decisions.
A retention number is a definition before it is a figure. Change what counts as retained and watch the answer move — that argument is the actual work.
Work, study, and cities on one axis.
The overlap is the information — where the threads run at the same time.
Here's what I'd build for you.
Scoped pieces of work with timelines that are what they actually took, not what they should have taken.