Chatbot eval creation

The problem

A recipe chatbot can sound confident and still get things wrong. The question was how to find those failures systematically and measure them, instead of spot-checking a few answers.

My role

Coursework for the Maven AI Evals course. I did it solo, from writing the bot’s prompt to analyzing where it failed.

What I built

• A system prompt defining the bot’s role, rules, safety behavior, and output format
• A set of test queries, then error analysis of the bot’s answers: reading the traces and grouping what went wrong into a taxonomy of failure modes
• An LLM-as-judge to score the bot’s answers automatically
• An analysis of conversation traces showing where a multi-step cooking agent breaks down

Outcome

The clearest failure: the bot dropped a dietary restriction when the user implied it instead of stating it.

My LLM judge checked whether each recipe actually met the user’s stated dietary restriction. It agreed with my own labels most of the time; its misses were false negatives, marking a recipe as failing when it was actually fine. I deliberately kept the judge to stated restrictions, where pass or fail is clear-cut. Implied restrictions are the harder problem, and they’d need their own check.

In the multi-step agent, most failures came from skipping the argument-building step.

What I took away:

• Read real traces before writing any eval.
• Don’t trust a judge until you’ve checked it against labels you created yourself.