ML systems and evaluation
I take a model from a demo to something that runs every day. Then I test whether it does what it says.
- Production ML at Booking.com: promo targeting, booking-failure prediction, and ML pipelines in Flights.
- A 44-task test suite for agents. 22 models have been through it.
- A build system where agents write the code, so I can see how they fail.
See it: Agentosaurus · agent test suite · BriefGPT