// certification_earned
AWS Cloud Quest: Machine Learning
Badge earned after completing all 25 quests across foundations, AI services, ML engineering and generative AI.
A walkthrough of every AWS Cloud Quest: Machine Learning assignment — the problem each one solves, the AWS services involved, and what it builds toward as a practising data engineer.
A city council needs a static website that survives traffic spikes at zero operational overhead — no servers to patch or provision.
// solution
Migrated to S3 static website hosting — globally available, infinitely scalable, zero infrastructure.
A new team needs secure, auditable AWS access without sharing root credentials or over-provisioning permissions to anyone.
// solution
Configured IAM users, groups and least-privilege policies — the security bedrock every subsequent quest builds on.
A startup needs backend logic that scales to zero during idle periods and costs nothing when not in use — no EC2, no idle bill.
// solution
Built an event-driven Lambda exposed via API Gateway — the serverless pattern underpinning every AI service integration later in the course.
An e-commerce platform has unpredictable traffic surges. Manual capacity planning means constant over-spend or downtime.
// solution
Deployed an Auto Scaling Group behind a load balancer — automatic right-sizing based on real demand, no human intervention needed.
A healthcare company needs to isolate sensitive workloads from public internet traffic while still allowing controlled external access.
// solution
Designed a VPC with public/private subnet separation, Internet Gateway and security group rules — network isolation done properly.
A finance team has zero visibility into cloud spend. Bills arrive as a monthly surprise with no attribution to teams or workloads.
// solution
Implemented tagging strategies, budget alerts and Cost Explorer dashboards — full spend attribution and early warning on overruns.
An application needs relational data for transactions and a NoSQL store for low-latency session data — two workloads, two different requirements.
// solution
Provisioned RDS for structured queries and DynamoDB for high-throughput key-value lookups — right tool, right job.
An organisation needs to prove data at rest is encrypted and every API action is auditable for compliance and governance.
// solution
Enabled KMS-managed encryption on S3 and RDS, CloudTrail capturing all API activity — encryption and auditability locked in.
A logistics company generates thousands of sensor events per second. Batch ETL is too slow — insights are needed in near real-time.
// solution
Built Kinesis → Firehose → S3 streaming pipeline — real-time data delivery for downstream ML and analytics.
A security firm needs to detect objects, faces, and unsafe content in thousands of images per day — without building a CV model from scratch.
// solution
Integrated Rekognition via Lambda — automated object detection, face analysis and content moderation at scale, zero model training.
An accessibility platform needs to convert written content into natural-sounding audio across multiple languages — without any audio recording infrastructure.
// solution
Used Polly with SSML markup to generate lifelike speech stored in S3 — fully managed, no audio engineering required.
A legal firm receives thousands of scanned contracts. Manually extracting key fields is slow, error-prone, and impossible to scale.
// solution
Used Textract to extract structured text, tables and key-value pairs from scanned PDFs — unstructured docs turned into queryable data.
A retail company wants to understand sentiment trends across thousands of customer reviews without reading them manually.
// solution
Ran Comprehend batch jobs to score sentiment and extract key phrases at scale — turning raw reviews into actionable signal.
A call centre needs to automatically convert recorded calls into searchable transcripts for quality assurance and compliance review.
// solution
Configured Transcribe to process audio from S3, outputting timestamped transcripts — enabling downstream NLP on real conversations.
A facility needs to detect unauthorised drones in surveillance footage — a niche object detection problem the generic Rekognition model doesn't cover.
// solution
Trained a Custom Labels model on drone imagery — fine-tuning AWS computer vision to a specific, real-world detection task.
A data science team wastes weeks provisioning GPU instances and managing Jupyter environments before any modelling can begin.
// solution
Spun up SageMaker Studio with managed notebooks and data connections — a full ML dev environment in minutes, not weeks.
An ops team needs to detect unusual spikes in business metrics before they become customer-impacting incidents — no ML expertise on the team.
// solution
Configured Lookout for Metrics on time-series data — automatic anomaly detection with root-cause groupings, no model training required.
A team has a custom ML model in Python that needs production-grade inference endpoints — without rewriting it for a specific framework.
// solution
Containerised the model with Docker, pushed to ECR, deployed on SageMaker as a real-time endpoint — BYOM done properly.
An enterprise's foundation model gives generic responses. They need domain-specific answers tuned on proprietary data without training from scratch.
// solution
Fine-tuning an LLM via JumpStart on domain-specific datasets — adapting foundation model behaviour at a fraction of full training cost.
A team needs to train and deploy a custom image classification model using TensorFlow, without managing the underlying GPU infrastructure themselves.
// solution
Used SageMaker's TensorFlow estimator to train a CV model on managed GPU instances and deploy it as a real-time inference endpoint.
An engineering team wants to train an agent to learn optimal decision-making through trial and error — without provisioning simulation environments manually.
// solution
Trained a reinforcement learning model on SageMaker using managed simulation environments — agent learning through reward-based feedback loops.
A product team wants LLM capabilities in their app but has no ML expertise and can't afford to train a model from scratch.
// solution
Used Bedrock to call foundation models via API — text generation, summarisation and Q&A, zero model infrastructure to manage.
A creative agency needs to generate product imagery on demand from text descriptions — cutting design turnaround from days to seconds.
// solution
Invoked Stable Diffusion via Bedrock with engineered prompts to generate imagery programmatically, stored directly to S3.
A customer support team needs an intelligent chatbot that understands natural language intent and gives contextually relevant answers — beyond rigid keyword matching.
// solution
Built a conversational chatbot backed by an LLM via Bedrock, with Lex handling dialogue management — context-aware responses at scale.
The final milestone — completing the full AWS Cloud Quest: Machine Learning role from start to finish, across all 25 quests.
// achievement_unlocked
All 25 quests completed. Full ML role badge earned. Built across nights, weekends and stolen hours between Jet2, the NHS and a degree.