Talia Ben-David

I get called when an agent system works in the demo and fails in the field, and nobody in the room can explain the gap.

How I work through it:

1. Build the evaluation suite first. You cannot fix what you are not measuring, and almost nobody arrives with one.
2. Reproduce the failures on real traffic, not on the happy path someone recorded for the board.
3. Rank the fixes by cost. Two prompt changes and a retrieval fix usually beat the retraining that was proposed.
4. Leave the harness behind, so the next regression is caught by your team rather than by a customer.

Four enterprise deployments remediated this way. Former research engineer, independent since 2023, one slot roughly every two months.