Play free

AWS · Domain

AI and machine learning

Build, train, secure and run AI on Amazon Bedrock and SageMaker AI. 5 worlds, 30 stages, about 12 hours to complete.

FeedsDEA-C01AIF-C01MLA-C02AIP-C01

AI World 1 · Associate

GenAI Foundations

Choose, size, tune, prove, guard and govern AI services on AWS, the way the exams ask.

6 stages~2.5 h1 free

  1. 1Right ToolFreePick the AWS AI service each requirement needs, and know why not the others.
  2. 2Pick the ModelModality, context window, latency, where the weights live, and how each call is paid for.
  3. 3Shape the AnswerTemperature, length, stop sequences, few-shot and chain-of-thought, and a prompt that cannot be hijacked.
  4. 4Teach and MeasureFine-tune, pre-train, distil or retrieve: match the method to the data, then pick the metric that proves it.
  5. 5Guard the GateBlock the jailbreak, deny the topic, mask the PII, ground the answer, enforce it from IAM, and keep the record.
  6. 6Govern ItRoles not keys, logs the auditors can read, a Model Card, a Clarify report, a human in the loop, and Macie on the bucket.

AI World 2 · Associate

Grounded Answers

Build, tune, search and prove a knowledge base, so answers come from your documents.

6 stages~2.5 h

  1. 1Feed the BasePoint the data source where the documents are, parse what is on the page, and sync, or nothing is found.
  2. 2Cut and EmbedChunks that keep their context, an embedding model that reads the language, and vectors no bigger than the budget.
  3. 3Pick the StoreOpenSearch Serverless, Aurora pgvector, Neptune Analytics or S3 Vectors: the store decides what a query can do and what it costs.
  4. 4Find the Right ChunkRetrieve or RetrieveAndGenerate, a metadata filter, a reranker and enough results: the query decides which chunk the model reads.
  5. 5Prove the RetrievalRetrieve-only or retrieve-and-generate, and the metric that answers the question: coverage, relevance, faithfulness, citations, correctness.
  6. 6Nothing FoundFive tickets, one symptom each: read the sync history and the retrieval trace before you touch the model.

AI World 3 · Associate

Agents and Tools

Give an agent tools, a home, a memory, a budget, a guardrail and a trace.

6 stages~2.5 h

  1. 1Give It ToolsAn agent without tools answers from the model. Give it the right target, with the credential its target takes.
  2. 2Run It SomewhereLambda stops at 15 minutes, a session idles out, a session ID has rules, and DEFAULT moves on every release.
  3. 3Remember MeA stateless function remembers nothing. Give the agent a memory with the strategy, namespace and expiry each requirement needs.
  4. 4Know When to StopBound the loop, enforce the rule at the gateway, scope three roles, and give each tool the time it needs.
  5. 5Guard the AgentA guardrail on the model sees the user and the answer. What comes back from a tool needs its own check.
  6. 6Read the TraceSix tickets about a fleet of agents. Read the trace before you touch the model, then measure the agents you fixed.

AI World 4 · Associate

ML Pipeline

Prepare the data, label it, train and tune, measure bias, read the numbers, wire the pipeline.

6 stages~2.5 h

  1. 1Shape the DataClean without code, one feature group for training and scoring, a quality score on the catalogue, a 2 TB job that starts now, and data in the job's Region.
  2. 2Label ItBoxes on 8,000 photos within budget, a 900-photo pilot, documents that stay in-house, notes in six categories, and a person on the low-confidence payouts.
  3. 3Train and TuneNo code for the classifier, the team's script unchanged, a 6-hour job at up to 90% less, a tuning budget spent wisely, last month's search reused, and runs you can compare.
  4. 4Check the BiasCount the under-25s, compare their labels, compare the model's rejections, explain one decline, and write the card the regulator reads.
  5. 5Read the NumbersA 2% default rate scored honestly, the misses counted, two models compared at every threshold, a 99%/71% gap closed, and an error in dollars.
  6. 6Wire the PipelineFive MLOps tickets: a weak model in production, an unapproved version live, an evaluation with no model, a 3-hour step rerun, and a retraining nobody can trace.

AI World 5 · Associate

Serve and Secure

Serve the model on the right endpoint, scale it, keep it private, lock it down, test it live, watch it drift.

6 stages~2.5 h

  1. 1Pick the EndpointLoan decisions in 100 ms, a 40 MB scan that takes 20 minutes, 30 million rows overnight, five calls an hour, a GPU model and 400 tenants: one endpoint type each.
  2. 2Scale ItLunchtime peaks, a tool idle overnight, a scan queue, a fleet that flaps, trees on a GPU and a Neuron-compiled model: the policy, metric and family for each.
  3. 3Keep It PrivateA loan app with no internet path calls its model and Amazon Bedrock through the VPC, only what it should, and a model container that never calls out.
  4. 4Lock It DownThe bank's key on the volumes and the results, one caller for one endpoint, an execution role that can load the model, and an audit trail of every invocation.
  5. 5Test in ProductionA new container with no user impact, a 90/10 split, QA on the new variant, a canary that rolls back, yesterday's model back in a minute, and a serverless model's candidates.
  6. 6Watch the DriftSix tickets from a month in production: drift nobody saw, accuracy that fell, a fairness rule, a classifier with no monitoring, a doubled token bill and the auditors' prompts.

Game types in this domain

How each one plays