Building a product solo with AI coding agents, writing code got cheap.
What didn't: reading every diff, running the tests myself, and saying no when the agent is confidently wrong.
Review is the job now.
CODEOWNERS gotcha: docs/* only matches files directly inside docs, not nested folders.
If you want everything under it, write docs/ instead.
My own CODEOWNERS parser got this wrong too.
A regex in my code reviewer hung for 11 minutes on a single PR.
The cause: minified JavaScript. Long lines full of escapes sent the pattern into catastrophic backtracking.
The fix was a scanner that reads each line once.
If an AI reviewer leaves 10 comments on your PR and 3 are wrong, do you still read the other 7?
Or does one bad comment teach you to ignore the bot for good?
All four stages are built and working end to end. Building this in public from here on: what ships, what breaks, what I find wrong with my own assumptions. Follow along.
Introducing Stratum, a living intelligence layer for your codebase, not just a PR bot. Most AI review tools read a diff, leave a comment, and forget the PR ever existed. Stratum doesn't stop there.
Drift: a live view of your architecture's health, computed from everything the Review stage has learned over time. Which modules are quietly accumulating risk, which boundaries are eroding.
Incident: when something breaks in production, correlates the live error (Sentry/Render/Railway) against your reviewed PRs to rank likely root causes. "Why is this broken" becomes a ranked list, not a guessing game.
Not affiliated with @GoldmanSachs, but I noticed a possible React hydration issue on the careers page.
Console shows minified React errors #418 and #423. Some blocked analytics logs are also present, likely extension-related.
Sharing in case useful for the web/frontend team.