now
- building
- lossproof — one CDC stream into two Iceberg lakes, faulted on purpose, then every row reconciled against the source; leakproof — the same fraud model built carelessly and point-in-time correct, both scored on features a live stream produced; nightshift — a receipt-extraction lake with an LLM as an ETL step, and an agent on call measured against a 101-line runbook; this site, still — a terminal you can query instead of a portfolio you scroll
- reading
- Designing Data-Intensive Applications — the stream processing chapters, again; the Iceberg and DuckDB docs, mostly to argue with myself about table formats
- learning
- AWS Certified Solutions Architect – Associate, in progress
Four projects since September have turned into one method, mostly by accident:
build the careless version on purpose, measure exactly what it costs, and only
then build the one that is correct. warehouse-gpt put a number on what context
is worth to a text-to-SQL agent, leakproof on what leakage is worth to a fraud
model, lossproof on what an ordinary week costs a change-data-capture lake, and
nightshift on what an on-call agent is worth next to a plain runbook.
The through-line is the same conviction four years of moving data left me with — the hard part of applied AI is not the model, it is getting the right data to the right place, on time, in a shape someone can trust. What has changed is that I would now rather count that than assert it, which is why each of these ships the number that argues against it too.