Case studies
What actually shipped.
Anonymized by request — real engagements, generic attribution. We show what broke and what we changed, not borrowed names or logos.
AI support chatbot · LLM behavior
The assistant regressed on every prompt change, with nothing to catch it. We built the eval suite that turned that into a gate.
no eval coveragegraded golden set on CI
Read the full story No-code e-commerce shop · checkout
Checkout silently dropped orders under a race condition no test covered. We reproduced it, fixed it, and gated it with a regression test.
orders lost, unnoticedrace fixed · regression test gating checkout
Read the full story Clinic booking app · scheduling
A timezone bug double-booked slots and overran the calendar. We corrected it across time zones and covered it with tests.
double-booked slotscorrect across time zones · covered by tests
Read the full story