Evals in AI: A Deep Dive — Tejas Kumar, IBM
- Source
- youtube.com
- Author
- AI Engineer
- Date
Context
A passing automated test can be wrong about whether an AI system behaves correctly, and Tejas Kumar builds a concrete case in an AI Engineer workshop. A customer support answer passes a check because it contains the word cannot, even though it approves a return outside the store's policy. Simple assertions and fuzzy matching cannot tell the difference.
The workshop's fix is an judge, a model that grades answers against a dataset of customer scenarios and human verdicts. According to the description, the judge itself needs checking: it starts favoring an answer from its own model family, and adding the real return policy improves its decisions but leaves gaps. Kumar names position bias, sycophancy, self preference and verbosity as ways a judge can mislead.
His process is to compare the judge with human verdicts, set an 80% agreement threshold as a continuous integration gate, and keep refreshing the dataset with fresh traffic and human review. He also recommends starting with deterministic checks before paying for model judgments. These are the speaker's recommendations as summarized in the description; the session itself was not available.
- Kumar's opening failure: a support reply passes a test because it contains the word "cannot", even though it approves a return outside store policy. Substring and fuzzy assertions stop being useful at that point.
- He names four ways an LLM judge can mislead: position bias, sycophancy, self preference and verbosity. In the demo the judge favors an answer from its own model family, so it must be against human verdicts.
- Supplying the actual return policy improved the judge's decisions, but missing details and a loyal customer's appeal still exposed gaps, and Kumar eventually changed the model rather than trusting a green test.
- He sets an 80% judge-human agreement threshold as a CI gate, and says fresh traffic plus human review must keep the scenario dataset current.
- His cost advice: start with deterministic checks before paying for model judgments, and treat pre-deployment and runtime protections as separate layers.
- LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.
- eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.
- calibration — How well a model's confidence matches reality — a calibrated model saying "90% sure" is right about 90% of the time.
- agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
Checking sign-in…
Loading comments…






