What we've learned shipping AI agents to production at growth-stage companies. No thought-leadership. Original data, customer-signed numbers, and the boring infra wins behind the headlines.
An independent verifier model trained on customer-corrected outputs catches 4× more bad outputs than self-consistency alone. The architecture, the eval, and what we got wrong twice.
Read article →What happens when you reprice your product against the labor it replaces. Six months of usage data, three painful contract resets, one obvious conclusion.
Read article →The browser extension that captures cross-app workflows. Presidio + a custom classifier. Why we open-sourced the client and what threats we found.
Read article →Golden datasets, regression nets, shadow runs in production. Why our PRs block on accuracy, not on test pass rate.
Read article →The patterns we saw, the ones we expected and didn't see, and the single workflow that pays for the year.
Read article →The visual system behind Routix. Editorial typography, Swiss grid, one signal color — and the deliberate absence of gradients and glow.
Read article →