Hands-on learning path · Cloud

Reading builds vocabulary; building builds fluency. These eight stages take roughly three to four months part-time and around €200–500 in cloud spend, if you tear everything down after each stage. Each stage reuses the previous one, and the final stage is a portfolio-grade system.

Two rules for every stage: set a budget alert before launching anything, and before each step write down what you predict (memory, speed, cost) using Part 5, then compare with what you measure. The gap between prediction and measurement is where the learning happens.

Stage 0 - Foundations (1–2 weeks, free)

Build a FastAPI service that serves a small embedding model on CPU, and package it as a Docker image. Learn the Linux shell, SSH, processes and systemd, file permissions, ports, curl and dig, Git, and Dockerfiles.

docker run serves an /embed endpoint, and you can explain every line of the Dockerfile.

Stage 1 - One machine by hand (1 week, a few euros)

Create a cloud account with multi-factor login and a budget alert, and never use the root account for daily work. Launch a small VM in Frankfurt, SSH in, run your container, put Caddy or nginx in front with a TLS certificate on a domain, and lock the firewall down to ports 22 and 443. Then destroy it. A cheap Hetzner server works just as well for this stage.

You can draw the full path of a request from a browser to your container, naming every hop.

Stage 2 - Serverless, object storage and identity (1 week, a few euros)

Deploy the same container to Cloud Run, or to Lambda or Fargate. Build an event pipeline: a document uploaded to a bucket triggers a function that extracts text, embeds it and stores it in Postgres with pgvector. Use roles for every permission and a secrets manager for every secret.

The pipeline works end to end with no long-lived keys anywhere.

Stage 3 - Your first GPU (1–2 weeks, about €20–50)

Rent one GPU (an L4, L40S or A100 on a marketplace or serverless GPU platform) and serve an 8B open model with vLLM. Load-test it at 1, 8, 32 and 128 concurrent requests, measuring time to first token, time per token and total throughput. Compute your cost per million tokens and compare it with a hosted API’s price.

You can explain the shape of the throughput-versus-concurrency curve using memory bandwidth and the KV cache.

Stage 4 - Fine-tune and recover (1–2 weeks, about €20–80)

Run a LoRA or QLoRA fine-tune of a small model on a dataset stored, versioned, in object storage. Track the run in Weights & Biases or MLflow and checkpoint to object storage. Deliberately kill the instance mid-run and resume on a new one. Evaluate before and after on a held-out set, then serve the adapter with vLLM.

Resume from checkpoint works, and the evaluation shows a measured difference.

Stage 5 - Infrastructure as code and CI/CD (1–2 weeks, a few euros)

Rebuild Stages 1 and 2 in Terraform or OpenTofu with remote state. Add a GitHub Actions pipeline that authenticates through OIDC, builds and pushes the image, runs a small evaluation suite, and deploys only if it passes.

destroy followed by apply recreates the whole system with no manual steps.

Stage 6 - Kubernetes and multiple GPUs (2–3 weeks, about €50–200)

Learn Kubernetes locally first with kind or k3d. Then create a managed cluster (GKE is the gentlest start) with a GPU node pool, deploy vLLM, autoscale on queue length, and time a cold start. Separately, run an FSDP training job across 2–8 GPUs on one node with torchrun, and optionally a two-node run launched through SkyPilot.

You can explain your measured scaling efficiency from 1 to N GPUs, and where the cold-start time goes.

Stage 7 - Capstone: a production-grade AI system (3–4 weeks)

Build a RAG or agent system end to end, deployed in an EU region:

Ingestion through a queue and workers, with idempotent processing.
Hybrid retrieval with reranking.
A model gateway with a fallback provider.
Tracing with Langfuse or similar, and dashboards for time to first token, cost per request and quality.
An evaluation suite in CI that gates deploys.
Budget alerts, and a pass through the security checklist in Part 16.

Finish with a short write-up: an architecture diagram, measured latency, cost per request, and what you would change at ten times the load.

The write-up explains every design choice with a number or a constraint.

Certifications

They are optional. AWS Solutions Architect Associate, Google Professional Cloud Architect or Machine Learning Engineer, and the Certified Kubernetes Administrator (CKA) give a structured syllabus and a recognizable line on a CV. Built and measured projects carry more weight in interviews for AI infrastructure work.