You're deploying the recommendation model to the cloud for a Berlin-based fashion marketplace -- multi-service Docker Compose orchestration, evaluation-gated deployment, cost controls, secrets management, and scaling configuration within a fixed monthly budget.
The discipline skills: orchestrating three Docker Compose services that must communicate correctly (not just start), deploying to AWS with budget alerts and auto-shutdown, managing secrets in a cloud environment (not local .env files), configuring scaling as a cost-responsiveness trade-off, and linking model artifacts to their training data and configuration for full provenance.
The AI-direction lesson: cloud infrastructure introduces dimensions where AI's defaults cost real money. AI generates Docker Compose files where services use localhost to reach each other instead of Docker service names, causing silent connection failures. It deploys to cloud without budget alerts, auto-shutdown, or cost controls. It puts API keys in plaintext. It creates CI/CD pipelines that deploy without running the evaluation suite. Each of these defaults is expensive or dangerous. The skill is writing explicit constraints before AI generates any infrastructure code -- "pin all dependency versions," "eval gate must exit 1 on failure," "never store secrets in workflow YAML" -- because on cloud terrain, catching a bad default after deployment means real cost and real exposure.
Your Role
The brief is familiar territory -- deliberate ambiguity, no guides, you own the decomposition. What changes is the terrain. Cloud deployment introduces real money: a service left running overnight costs real euros. Secrets management introduces real security exposure. And orchestrating three services that must communicate correctly is a different problem from running a single container.
Marcus Webb appears as your senior colleague. He has opinions about scaling strategies and cost controls.
What's New
Last time you built a versioned feature pipeline with DVC and a monitoring dashboard. The model existed in your development environment.
This time the model goes to the real world. Docker Compose enters for multi-service orchestration -- three services that need to find and talk to each other, not just start. Cloud deployment introduces financial reality: every minute a service runs costs money, and your first task is figuring out how much. Evaluation gates, which you built in CI/CD two projects ago, now stand between your code and a cloud deployment where a bad model costs real money to serve.
The hard part is not any single deployment step. It is that cloud infrastructure introduces dimensions -- cost, security, coordination -- that your local environment completely hid from you.
Tools
- Docker Compose -- multi-service orchestration (new)
- AWS CLI / boto3 -- cloud deployment (new)
- W&B -- experiment tracking alternative, discussed but not adopted (new)
- Python -- scripting, deployment configuration (familiar)
- FastAPI -- serving endpoint (familiar)
- MLflow -- experiment tracking (familiar)
- DVC -- data/feature versioning (familiar)
- scikit-learn -- modeling, evaluation (familiar)
- Git / GitHub -- version control (familiar)
- Claude Code -- AI direction, plan mode (familiar)
Materials
You receive:
- The recommendation model from the previous project (serialized model and feature configuration)
- A FastAPI serving application with health check and prediction endpoints
- An evaluation suite that gates deployment (model must pass before deploying)
- A Docker Compose template with three services -- deliberately incomplete
- A cloud deployment configuration template with TODOs for secrets, scaling, and cost controls
- Infrastructure requirements from Ravi (the backend developer)
- A project governance file (CLAUDE.md)