Big build day, almost all of it on A Wine App — roughly 7.5k lines added across 26 commits.
The morning started with paper, not code. I tested how well an LLM could generate documentation for the wine database, wrote up a strategy for splitting intelligence between local and online models — then superseded that strategy the same day after a second review pass. The revised plan is what actually got built: an evidence-bundle layer that pulls together what the app knows about a wine — producer, region, grape — plus a deterministic local explanation engine on top, fixtures, reports, a request composer, and contracts for when an online model is allowed to assist. The boundary between “the app explains from evidence” and “a model fills in prose” is now documented and validated, not vibes.
In between, the running app kept me honest. Reviewing it with a fresh zero-rating profile surfaced the worst kind of bug: the sommelier confidently claimed I’d loved a Barolo I’d never rated. Fixed it to write honest prose when there are no ratings and regenerate on every new one. Also chased down catalog data showing grapes as raw stringified lists (about 1% of wines, fixed at the scraper source), made the grape filter from the profile screen return real results instead of nothing, and taught the debug seeder to write explicit ratings so testing with real feedback data is actually possible.
Two bits of hygiene: a CI gate now blocks new hardcoded user-facing strings in views, and I moved the backend off the cloud database onto local Postgres — one less moving part while the app is still pre-launch.
Also jotted a single line into the ideas file. That’s all it gets.
Tonight: the intelligence engine has a real foundation and tested boundaries; next is wiring it into what users actually see.