Come learn how
QA meets AI.
Structured lessons, deep-dive articles and a podcast: for testing software with AI, and testing AI itself.
Setting Up a Movie Dataset You Can Actually Trust
Before an LLM agent can recommend a single movie, the dataset underneath it needs to be verifiably correct. A pandas loader, an exploration pass, and the assertions that catch a bad dataset before it reaches the agent.
Read itDeep-dives with runnable code
Long-form engineering write-ups: LLM evals, regression harnesses, self-healing suites. Every post ships with code you can run.
Browse articlesVideos, a podcast and guided paths
Watch, listen, and build. Structured lessons that take you from testing basics to shipping LLM evals in CI.
Start learningHands-on help for your team
Consulting, eval setup and workshops from an engineer who does this in production, not in slides.
See servicesFresh from the lab.
Setting Up a Movie Dataset You Can Actually Trust
Before an LLM agent can recommend a single movie, the dataset underneath it needs to be verifiably correct. A pandas loader, an exploration pass, and the assertions that catch a bad dataset before it reaches the agent.
A Naive Chatbot Is Not an Agent
Phase two of the movie recommendation project is called an agent. It has conversation memory and a system prompt, and it cannot do a single thing besides generate text. What actually separates a chatbot from an agent, and how do you test something that can only talk?
Giving the Agent Tools It Can Use
Four plain Python functions turn a dataset the model can't see into one it can query: search, and filter by genre, language, and rating. None of it touches an LLM yet, which is exactly why it's the easiest, highest-value code in the whole project to test.
LLM evaluation, demystified
Ground-truth sets, judges, thresholds: how to score what can’t be asserted.
Automation that maintains itself
Self-healing suites, AI-assisted test generation, and where to draw the line.
CI patterns that scale
Nightly drift checks, model-upgrade gates, and builds that fail for good reasons.
No hype, receipts included
Every claim comes with code, honest limitations, and a way to reproduce it.