60 days · 4 phases · 31 graded labs · 5 design briefs · $200
AI Platform Engineering Architecture
Build the platform an AI team actually runs on.
Four phases: the cloud-native baseline, infrastructure as code and orchestration, the GPU and network layer an inference platform needs, and whole-system design. Phases 1–2 are hands-on from day one; Phase 3 grades the manifests and policies you would apply to real accelerators; Phase 4 is design work against a rubric.
What runs in the sandbox
- Linux, Git, networking, Docker and shell/Python labs — graded on real filesystem state.
- Terraform and Kubernetes: parsed configuration and cluster objects, graded on state.
- AWS compute, IAM and S3: the CLI writes a simulated account that the checks read back.
- GPU serving manifests: image, model argument, accelerator request and Service.
- The CI/CD workflow file itself — triggers, steps and what it applies.
What it cannot do
- No GPU is provisioned. The scheduling request is graded; CUDA and the GPU operator are reading.
- No VM boots, and nothing is pushed to a real registry or cloud account.
- FSx for Lustre, EFA/RDMA and DCGM metrics are not simulated; the labs grade the policy around them.
- No model weights are downloaded, so “serving” means the manifest is correct, not that tokens stream.
- Phase 4 is not graded. The rubric is the check, and the drawing is yours.
Foundations & the control plane
Days 1–20
The cloud-native baseline every later phase assumes.
Compute: Linux, Docker and cloud sandboxes. The milestone registers a VM rather than booting one.
| Days | Topic | Environment | Prove it |
|---|---|---|---|
| Days 1–3 | Pick an SDLC process, including the agentic variantA decision, not a keyboard skill — there is no graded lab. Write the process down; the criterion is that a new engineer could follow it. | Reading + a written decision | No graded lab |
| Days 4–5 | Operating systems and Linux: files, permissions, processes | Linux sandbox | |
| Days 6–7 | Networking: TCP/IP, DNS, HTTP, load balancing | Linux sandbox | |
| Days 8–9 | Source control: branching, merging, pull requests | Git sandbox | |
| Days 10–11 | Scripting: Bash first, then Python | Linux sandbox | |
| Days 12–13 | Containers: images, containers, Docker Compose | Docker sandbox | |
| Days 14–16 | Core cloud services: compute, storage, IAM | Cloud sandbox | |
| Days 17–20 | Milestone: containerize an app and register it on a cloud VMThe sandbox grades the image and the instance/security-group record. It does not boot a VM or push the image to a registry. | Docker + cloud sandbox |
Core DevOps & orchestration
Days 21–40
Infrastructure as code, Kubernetes, pipelines and observability.
Compute: Terraform, Kubernetes and Git + terminal sandboxes. Nothing here needs a real cluster.
| Days | Topic | Environment | Prove it |
|---|---|---|---|
| Days 21–25 | Terraform: provisioning infrastructure as code | Terraform sandbox | |
| Days 26–31 | Kubernetes: pods, deployments, services, config | K8s sandbox | |
| Days 32–35 | CI/CD pipelines: GitHub Actions or GitLab CI | Git + terminal | |
| Days 36–38 | Monitoring and observability: Prometheus, Grafana, logs | K8s sandbox | |
| Days 39–40 | Milestone: automated CI/CD deployment to Kubernetes with TerraformOne lab spans all three: a namespace from Terraform, the workload in the cluster, and the workflow that redeploys on a tag. The runner is simulated — no job executes. | Terraform + K8s + workflow file |
AI infrastructure & specialised hardware
Days 41–50
The accelerator, storage and network layer an inference platform needs.
Compute: K8s and cloud sandboxes. No GPU is provisioned: these labs grade the scheduling request, the manifest, the storage policy and the network boundary.
| Days | Topic | Environment | Prove it |
|---|---|---|---|
| Days 41–42 | GPU computing basics: CUDA architecture, NVIDIA GPU OperatorNo accelerator runs in the browser. The lab grades the GPU request a scheduler acts on; CUDA and the operator are reading. | K8s sandbox | |
| Days 43–44 | AI storage and data gravity: FSx for Lustre, MinIOS3 stands in for the object store. FSx/Lustre is reading — there is no parallel filesystem to mount. | Cloud sandbox | |
| Days 45–46 | AI networking: EFA, VPC topologiesEFAs and RDMA cannot be simulated. The lab grades the security-group boundary you would put around a training fleet. | Cloud sandbox | |
| Days 47–48 | Model serving and inference: vLLM, TritonThe Deployment, the model argument and the Service are graded. No weights are downloaded and no request is served. | K8s sandbox | |
| Day 49 | AI observability: DCGM exporter, TTFT, throughputThere are no GPU metrics here. The lab teaches the scrape-and-alert shape that the DCGM exporter feeds. | K8s sandbox | |
| Days 49–50 | AI security: model weight protection, secrets management | Cloud + K8s sandbox | |
| Day 50 | Milestone: serve an open-weight model on Kubernetes with vLLMThe two serving labs together are the milestone: request the GPU, then expose the OpenAI-compatible server. | K8s sandbox |
System design & whole systems
Days 51–60
Connecting the components at scale, on a whiteboard.
Compute: None. Design work, reviewed against a written rubric.
| Days | Topic | Environment | Prove it |
|---|---|---|---|
| Days 51–53 | Core concepts: scalability, load balancing, caching, databases, queues | Design brief | |
| Days 54–55 | Design a URL shortener | Design brief | |
| Days 56–57 | Design a notification system | Design brief | |
| Days 58–59 | Design a file storage service | Design brief | |
| Day 60 | Milestone: design an end-to-end LLM inference platform | Design brief |
How to work through it
- Read the row's topic, then open the lab before reading anything else.
- Finish every activity. A lab is done when the checks pass against the sandbox state, not when the commands look right.
- Write down what the row cannot prove — for the GPU and network rows, that note is the interview answer.
- Phase 4: draw first, then check the drawing against the rubric, then redraw.
The syllabus is a map, not a certificate: nothing here is timed, and no row is signed off for you. A lab's green ticks are the only claim the product makes.