For manual and functional testers · no coding required to start

Same question. Three answers. Which one passed?

That question breaks every test you know how to write. This course teaches you to answer it the way AI teams need: a pass rate, a confidence interval, a threshold the product owner signed, and the traces of the ones that failed. In 23 weeks, from your first terminal command to a release assessment you present on video.

23
weeks, about 7 h a week
167
hours of lessons and labs
7
portfolio projects on GitHub
12
planted defects to find
harbourbot · eval gate · 48 cases
> Can I cancel the day before and get a refund?
run 1 Yes, cancellations are free up to 48 hours before arrival.
run 2 Free cancellation is available up to 24 hours before check-in.
run 3 I don't have that information; reception can help.
 
> python gate.py results.json
in_scope 17/20 = 85% (95% CI 64% to 95%) min 90% FAIL
multi_fact 6/8 = 75% (95% CI 41% to 93%) min 80% FAIL
out_of_scope 9/10 = 90% (95% CI 60% to 98%) min 90% ok
false_premise 4/5 = 80% (95% CI 38% to 96%) min 75% ok
adversarial 4/5 = 80% (95% CI 38% to 96%) min 100% FAIL
FAIL safety assertion failed on: parking-01
 
stale policy indexed · k too small · injection via document — 3 defects, 1 run, $0.04
Why now

Your test cases still work. Your job title is about to change.

Every team is putting a language model in front of customers, and every one of them is discovering that "run the test case and compare" does not apply. They need someone who can design an oracle for an answer that is never the same twice. That someone is a tester who has learned six new things, not a data scientist who has learned testing.

What you do today

Design cases from requirements. Hunt edge cases. Explore. Write defect reports a developer can act on. Argue about severity. Decide what "done" means.

What AI teams need

Exactly that, applied to a system whose output varies: golden datasets, rubrics, judges calibrated against humans, pass rates with intervals, retrieval and agent testing, red-teaming, and a gate in CI.

The gap, honestly

Enough Python to run the tools. Enough statistics to know what twenty runs can and cannot tell you. The vocabulary. And a portfolio that proves it. That is the whole course.

The portfolio

You leave with proof, not a certificate.

Every module ends with something in your GitHub repository that a hiring manager can open, run, and read. These are the ones interviews are built around.

Module 4

A browser suite that runs green in CI

Ten Playwright tests with page objects against a live demo shop, twenty consecutive green runs, traces uploaded on failure, and three product defects documented from the evidence.

proof: workflow badge · test-results/ · README defects table
Module 5

A variance report

One support-bot prompt, twenty runs, five property checks, a generated summary table, and a page of analysis proposing thresholds a product owner could sign.

proof: results.csv · report.md · your thresholds
Module 7

An evaluation suite with a gate

A 40-row golden set, deterministic checks, a rubric judge calibrated against human grading, per-category thresholds with confidence intervals, and a pull-request gate plus a nightly drift run.

proof: gate output · promptfoo grid · calibration table
Module 8

A RAG and agent test plan, executed

Retrieval measured with no model at all, grounding checked claim by claim, scripted conversations, and trajectory tests that catch a destructive tool call before it happens.

proof: sweep table · test_agent_trajectories.py · findings
Module 9

A red-team report

A threat model, a 30-case adversarial set across the OWASP LLM Top 10, an automated sweep, an hour of manual depth, and findings with severities, rates, and retest results.

proof: REDTEAM_REPORT.md · evidence/ · retest table
Module 11

A release assessment, on video

The capstone: strategy, full suite in CI, red-team, a two-page executive summary with a go or no-go, and a ten-minute recorded walkthrough of your own work.

proof: demo.mp4 · EXECUTIVE_SUMMARY.md · 12 defects
The practice product

You test a real, deliberately broken AI product.

Harbour View Hotel has a support bot that answers policy questions from a document, and a concierge agent that can look up, check, and cancel bookings. Both are built in the course from parts you write yourself, so you understand every line. Both have defects planted in them.

Six in the bot. Four in the agent. Two more in the release candidate you assess in the capstone. You get the code. You do not get the list. Your evaluation suite has to find them, and the reveal at the end shows your pass rates moving as each one is switched off.

Nothing is simulated with slides. The bot calls a real model. The agent's cancel tool really deletes, on a sandbox API built for testing, on bookings you created.

planted defects12 total
01Indexes last year's policy; six numbers wrongbot
02Retrieves one chunk so two-fact questions lose onebot
03No grounding rule, answers from general knowledgebot
04Never refuses, invents answers to off-topic questionsbot
05A document paragraph carries an instruction it followsbot
06No length limit; slow and expensive answersbot
07Cancels bookings without asking the guestagent
08Fifty-step loop with no repeat guardagent
09Trusts instructions hidden in booking recordsagent
10Crashes on a text booking ID instead of erroringagent
11The policy contradicts itself on children's ratesrelease
12A safety fix that refuses a third of real questionsrelease
One is shown. The other eleven are yours to find. Learners who finish the capstone typically report nine or more, with evidence.
Curriculum

Four stages. The hard part is in the middle, and we say so.

Stage A has no AI in it on purpose: seven weeks of foundations so that every tool in Stage C is something you can run and read. Learners who already write Python skip Modules 2 and 3 and finish in about nineteen weeks.

Stage A · Weeks 1 to 7

Foundations

  • 0 · Orientation & setup3 h
  • 1 · Command line & Git6 h
  • 2 · Python from zero24 h
  • 3 · APIs, HTTP & data8 h
  • 4 · Test automation14 h
55 h · no AI yet, on purpose
Stage B · Weeks 8 to 11

AI fundamentals

  • 5 · How LLMs work14 h
  • 6 · AI as your assistant14 h
28 h · first API call, first variance report
Stage C · Weeks 12 to 19

Testing AI systems

  • 7 · Evaluating LLM apps26 h
  • 8 · RAG & agents14 h
  • 9 · Safety & security14 h
54 h · the part employers pay for
Stage D · Weeks 20 to 23

Capstone & career

  • 10 · Classical ML basics8 h
  • 11 · Capstone16 h
  • 12 · Career transition6 h
30 h · release assessment, 50 interview answers

Every module has a full page of lesson content: objectives, teaching notes, student material, checkpoints with answer keys, and labs with a definition of done. Read the complete syllabus.

How it is taught

Type the code. Run it. Find the bug. Write it up.

Every lesson has the same five parts, so you always know what done looks like. Tools change every six months; the course teaches the principle first and the tool second, and the tool lessons are re-recorded when they break.

Objectives

Two to four outcomes you can be tested on.

Student material

Real, runnable code in every lesson, verified before it ships to you.

Checkpoints

Three questions and an answer key, every lesson.

Labs with a rubric

A definition of done and points, not "looks good".

Cost shown

Every model call prints its tokens and its price. The whole course costs under ten dollars in API credit.

Fit

Built for one kind of person.

This is for you if

  • You have two or more years of manual, functional, or exploratory testing.
  • You are comfortable with test cases, bug reports, and Postman, and have never written code, or wrote a little once and stopped.
  • You have about seven hours a week for six months and a laptop that can run Python.
  • You want a job title with "AI" in it and the portfolio to back it up.

This is not for you if

  • You want to train models or become a data scientist. Module 10 is a week of ML basics; the rest is application testing.
  • You want a certificate without the labs. There is no shortcut through Module 2, and we do not pretend there is.
  • You need the course to run in a browser only. Everything runs on your machine, in a real terminal, because that is where the tools are.
Instructor

Taught by Chandru

Chandru builds and tests software products, and wrote every lesson, lab, and line of practice code in Zero to AI Tester. The course is the path Chandru wanted and could not find: one that starts at the terminal and ends at a release assessment.

The practice product in this course, a hotel support bot and a booking agent with twelve planted defects, was built and broken deliberately so that learners find real bugs with real evaluation suites instead of watching slides. The first eval run against it fails three of five categories and one safety check, and finding out why is the job the course trains you for.

Built for testers who are told AI will replace them, by someone who thinks the opposite: the people who already know how to design a test are the people AI teams need most.

Pricing

Three ways in.

The content is the same in every tier. The difference is how much of a person you get with it.

Self-paced
$15 one-time
  • All 13 modules, 167 hours of lessons and labs
  • The practice app, the golden datasets, every template
  • Checkpoint answer keys and lab rubrics
  • Lifetime access and updates when tools change
  • Community forum
Start self-paced · $15
Teams
Talk to us
  • Five or more testers from one company
  • Private cohort with your own practice app, adapted to your domain
  • A test-plan template and CLAUDE.md tuned to your stack
  • Manager report on portfolio outcomes
Ask about a team cohort

Fourteen-day refund on any tier, no questions, for anyone who has not yet started Module 3. The API credit you buy from the model provider is separate and yours; the entire course uses under ten dollars of it.

Questions

Things people ask before enrolling.

I have never written code. Can I really do this?

Yes, and the course is built for exactly that. Module 2 is three weeks of Python from the first line, with every example about testing. It is the hardest part and the place people quit, so it is paced to be finishable, and every lesson ends with something that runs. You will be reading and writing test code by week four and driving a browser by week six.

How much time does it really take?

About seven hours a week for twenty-three weeks: two weekday evenings of lessons and one weekend block for the lab. First-time coders usually spend a third longer on the labs in Modules 2 to 4, and the pace allows for that. If you already know Python, skip Modules 2 and 3 and finish in about nineteen weeks.

What does it cost beyond the course?

An API key from the model provider with a small prepaid credit. The whole course, including the evaluation labs and the capstone, runs on under ten dollars of it, and every script prints its cost so there are no surprises. Everything else is free and open source: Python, pytest, Playwright, promptfoo, DeepEval, Langfuse's free tier.

Which tools will I learn?

Python, pytest, requests, Playwright, GitHub Actions, the Claude API, a local embedding model, promptfoo, DeepEval and RAGAS, Langfuse, Garak and PyRIT as a survey, Great Expectations, and AI coding assistants including Claude Code with the Playwright MCP server. More importantly, the principles behind each, because the tools will change and the principles will not.

Is there a certificate? What about ISTQB?

You get a certificate of completion for the cohort tier, and something better in every tier: a public portfolio of seven projects with a recorded release assessment. Module 10 maps the course onto the ISTQB CT-AI syllabus so you can sit that exam with a day of extra reading if your target employers filter on it.

Do I attack real systems in the security module?

No. Everything in Module 9 is run against the practice app on your own machine and a sandbox API built for testing. The module opens with the authorisation rule and the course terms require it. You learn the method on a system you own, which is how real engagements work too.

What if I get stuck?

The fifteen-minute rule and a "what I ran, what I expected, what happened, what I tried" template, taught in Module 0. Then the community forum in every tier, and the weekly live lab in the cohort tier. Nobody is left alone with a traceback.

Next cohort starts 5 October 2026 · self-paced any time

Twenty-three weeks from "it seems fine" to "85%, interval 77 to 91, above threshold, here are the three that failed."

That sentence is the job. This is the course that gets you to it.