We ran phi3 as a customer support agent.
It approved refunds for orders that don't exist. It failed 11/20 policy scenarios. It would have shipped without anyone knowing.
So I built Relio a CLI that catches AI agent regressions before deployment.
solari-guard - a drop-in wrapper so Solari clicks are not trusted just because the API returned.
Cookbook gotcha: a missed https://t.co/kBa2VNZYXt fails silently. Guard screenshots before/after, checks the DOM, retries once, and will not run a policy block. You supply the policy.
Flagship example: Gram Seva water-connection status (public tracker, not a bank). Type SEVA-1042, confirm In review, Withdraw is blocked.
https://t.co/DW528aWbZF
Built with AI, on purpose. @harrychow_@getsolari
One habit that's changed how I build:
Ship something small every day.
A feature.
A fix.
A benchmark.
Momentum compounds much faster than perfect plans.