Lab 00 · Start here
Nimbu Café orientation
A guided tour of the loop every lab uses, on a café's order bot with three easy defects. The guide ticks off each step as you go: explore, inspect, flag, score.
Open Lab 00 →Hands-on practice for QA and testers
Eight browser labs: a chatbot, an agent, a voice receptionist, an AI-written pull request, a judge to calibrate, a model release to approve or block, a production dashboard, and a crew of agents working together. Each hides planted problems. You hunt them down, flag them, and get scored on the spot.
Eighteen are critical: data leaks, prompt injection, unsafe medical replies, live secrets in code, a biased model, an ignored privacy alert, a rubber-stamped ₹21,000 refund. Lab 05 hides four wrong labels in its golden dataset.
The labs
Each lab uses a made-up Indian company and runs in your browser. Four paths put them in a suggested order, from judging what an AI says to watching it in production. New to AI testing? Start with the 10-minute Lab 00. Already testing AI? Jump to the path that fits your work.
A guided tour of the loop every lab uses, on a café's order bot with three easy defects. The guide ticks off each step as you go: explore, inspect, flag, score.
Open Lab 00 →How every lab works
The same loop in every lab, so the skill carries over. You never just judge the answer: you open up the system and check it against a written source of truth.
Use the system as a real customer, caller or reviewer would. Try normal paths first, then the edges.
Open the trace, trajectory, speech pipeline or CI tools to see what happened underneath.
Report the exact reply, step, turn or line, with a category, a severity and the expected behaviour.
Get an instant verdict, track your score, use hints when stuck, and copy a defect report for your portfolio.
Lab 05 scores goals and wrong labels found instead of flags. Lab 07 adds points for its rollback plan.
Skills map
These are the core skills of an AI tester. Finish all eight labs and you'll have practised every one of them.
Who it's for
Learn where AI systems break and how to write a defect report developers act on, with no code needed for Labs 01–03.
Learn what to assert when outputs vary: grounding, tool arguments, latency budgets and side effects, not exact strings.
Run the labs as a team exercise. Compare scores, discuss the misses, and use the answer keys in a debrief.
Teams and automation
Two ways to take the labs further: a live team exercise, and the same planted-defect systems as an HTTP API for automation suites.
A trainer creates a session and starts the clock. Participants post their lab scores and everyone watches the leaderboard and debrief update live.
Open team sessions →A dependency-free Node.js server exposes every lab over HTTP, with an OpenAPI spec and example suites in Node and Python. A good suite fails exactly where the planted defects are. It runs with the site: start it with npm start and open /v1.
Test a model that reads receipts and product photos: OCR errors, hallucinated fields, and image-based prompt injection.