By playing, you agree to uListen's Terms and Privacy Policy and YouTube's Terms of Service.
Summaries, chapters, transcripts, and FAQs on this page are made by uListen, not YouTube.
At a glance
WHAT IT’S REALLY ABOUT
Why AI Evals Matter and How to Build Your First
- Reganti argues that building an agent is easy but making it hold up in production is hard, which makes evals the key skill for AI builders.
- AI products differ from traditional software because user inputs are free-form and model outputs are probabilistic, and failures often appear as confident, wrong answers rather than errors.
- Model evals such as MMLU Pro and SWE-bench compare general model capability, but product evals are needed to test a system in your own environment, policies, and user base.
- The three ways to build evals are human evaluation as ground truth, deterministic code-based checks, and LLM judges that apply rubrics to subjective criteria.
- A walkthrough in Arize builds an escalation-check judge for a QuickCart support agent using a reference dataset, error analysis, a rubric, and a judge run, with the point that evals must keep being calibrated against real production data.
IDEAS WORTH REMEMBERING
5 ideasAI products are harder to test because both input and process are non-deterministic.
Users type free-form language rather than clicking constrained buttons, and the model generating outputs is probabilistic, so you can't enumerate and pre-test every case.
AI failures are often silent.
Traditional software throws exceptions and stack traces, while an AI system may return a confident, well-written, incorrect answer, so problems are often inferred from customer reactions.
Model benchmarks are a filter, not a verdict.
Benchmarks measure generic capability under standardized conditions, so they help narrow down models, but you must run product evals in your own environment with your own policies and users.
Each eval type has a distinct role.
Human judgment sets ground truth but doesn't scale, code-based checks handle objective pass/fail or similarity criteria, and LLM judges handle subjective, context-dependent judgments like politeness.
Error analysis tells you which eval to build.
By comparing AI answers to reference answers in the dataset, Reganti spotted that the agent was failing to escalate cases like double charges and account takeovers, which led to an escalation eval.
WORDS WORTH SAVING
5 quotesBuilding an agent is honestly the easiest part of building AI systems, and these days it looks like anyone can get something working in less than a day. So the hard part is making sure that it actually holds up in production.
— Aishwarya Reganti
when an AI system gets something wrong, it often just returns a confident, well-written, and completely incorrect answer.
— Aishwarya Reganti
evals are not a build once and forget kind of a thing.
— Aishwarya Reganti
So you simulate behavior before production, you build evals based on that behavior, and then you observe real production behavior and keep improving your evals until you actually trust them. That's the whole evaluation loop.
— Aishwarya Reganti
you need to be using your model evals just as a filter in order to make choices, but not entirely rely on them when you're building your products.
— Aishwarya Reganti
SPEAKERS
Aishwarya Reganti
hostAI engineer and educator focused on building and evaluating production AI systems and agents.
High quality AI-generated summary created from speaker-labeled transcript.
