
I Failed My AWS Beta Exam Instantly. So I Built My Own
After failing the AWS GenAI Developer beta exam, I rebuilt my 1,000-question quiz app from scratch, with a Strands Agent coach on Bedrock AgentCore Runtime
Series: AWS Quiz (2 articles)
- 1I Failed My AWS Beta Exam Instantly. So I Built My Own This article
I failed the AWS Certified Generative AI Developer – Professional beta exam. So I rebuilt my quiz app from scratch.
On December 16, 2025 — day one of the beta window for AWS Certified Generative AI Developer – Professional (AIP-C01) — I logged in at 9:00 AM sharp and took the exam.
I failed. Right there, on the spot.
This wasn't new territory for me. For the past few years I've half-jokingly called myself "the fastest man in the universe to fail a new AWS beta exam" — every time AWS drops a new certification, I sign up for the beta and go down in flames. It's become something of a personal tradition.
This one stung a little more, though. It's a certification that maps directly onto my day-to-day work with generative AI, and going into a beta exam means there's no study guide, no practice test bank, nothing except the official exam guide and whatever public information you can piece together. I studied on my own, walked in confident, and got humbled anyway.
Since April 2026 I've told myself "I'll retake it" more times than I can count, and every time, studying slipped down the priority list. I already had a quiz app I'd built for the first attempt, but it was thin — one format, no real exam pressure, nothing close to what the actual test feels like. Repeating the same approach was just going to produce the same result.
So I set a deadline: I'm retaking the exam next month. And instead of dusting off the old question bank, I tore it down and rebuilt it from zero. That rebuild is what this post is about.
What I built
A static web quiz app, styled like a Udemy-style course player, backed by a 1,000-question bank built directly from the official exam guide's domain weighting:
Domain | Focus | Official weight | Questions |
|---|---|---|---|
D1 | Fundamentals of Generative AI | 31% | 310 |
D2 | Designing Generative AI Applications | 26% | 260 |
D3 | Customizing Generative AI Models and Applications | 20% | 200 |
D4 | Operationalizing Generative AI Applications | 12% | 120 |
D5 | Applying Security, Compliance, and Responsible AI | 11% | 110 |
Total | | 100% | 1,000 |
Rather than one flat question format, I mixed in five answer-choice patterns that mirror what the real exam throws at you, so recognizing a memorized answer stops being enough:
- Single best answer (5 choices)
- Choose 3 correct answers (6 choices)
- Choose 2 correct answers (5–6 choices)
- Which one is incorrect (5 choices)
- Which two are incorrect (5–6 choices)
On top of the bank, there are four study modes:
- Random 200 — pulled from the full bank, explanation shown after every question
- Category 100 — pick one domain, drill it in isolation
- Mock exam, 85 questions — matches the real exam structure, with a 180-minute countdown timer. Scoring is withheld until the end, no per-question explanations, to reproduce actual exam pressure
- Flashcards, 150 questions — for reviewing on the go
An AI coach, not just a question bank
Solving questions doesn't tell you where you're actually weak — so I added a coaching layer built on a Strands Agent running on Amazon Bedrock AgentCore Runtime.
- Review my mistakes — reads your answer history and re-selects up to 30 questions, prioritizing what you got wrong or never attempted
- Analyze my weak points — aggregates your history and writes a short natural-language summary of which domains are dragging your score down and what to focus on next
The aggregation and selection logic is plain, deterministic Python — no LLM involved in the actual math. The Strands Agent (running Claude Sonnet 4.6) only reads those results and turns them into a coherent recommendation. Keeping the agent out of the arithmetic keeps the output consistent instead of occasionally hallucinating a percentage.

How it's wired together
The frontend is vanilla HTML/CSS/JS — no framework, no build step — hosted as a static site on Vercel. The AI coach path looks like this:
- Browser calls the "review mistakes" or "analyze weak points" action
- Amazon API Gateway (HTTP API) receives the request and triggers an AWS Lambda function
- Lambda calls the Strands Agent on Amazon Bedrock AgentCore Runtime via
InvokeAgentRuntime
- The agent calls Amazon Bedrock (Claude Sonnet 4.6) to produce the final response, which flows back through Lambda to the browser
Answer history lives entirely in the browser's
localStorage — there's no backend database. For a personal study tool, that's all the persistence it needs. (Architecture diagram attached separately.)Things that tripped me up
- AgentCore Runtime's Direct Code Deploy mode invokes
python agent.pywith no arguments. I had the entrypoint gated behind CLI subcommands, so the runtime just printed a usage message and never actually started serving. Fixed by defaulting to server mode whenever no subcommand is passed.
- A public Lambda Function URL came back
Forbiddendespite a correct resource policy — turned out to be an account-level guardrail blocking public Function URLs outright. Switched to an API Gateway HTTP API in front of the same Lambda and it worked immediately.

Where this leaves me
I still haven't passed the exam. But rebuilding the question bank from scratch forced me to re-read the exam guide's task statements line by line — and writing questions surfaces exactly where your own understanding is fuzzy in a way that passively reading never does.
I'm retaking the exam next month. This time I'd like to retire the "fastest man in the universe to fail" title for good.
Live app: https://aws-genai-pro-quiz.vercel.app
Source: https://github.com/yama3133/aws-genai-pro-quiz
Official exam guide: https://docs.aws.amazon.com/aws-certification/latest/ai-professional-01/
Series: AWS Quiz (2 articles)
- 1I Failed My AWS Beta Exam Instantly. So I Built My Own This article
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article