AI / ML Ops Learning Path
Where DevOps meets Machine Learning — build, deploy, and operate ML systems at scale. ~10–12 weeks at 5–7 hours/week, assuming basic programming knowledge.
By the end you'll be able to: containerize and serve a model behind an API, build a training pipeline with experiment tracking, run ML workloads on Kubernetes, and reason about LLM serving and inference costs — the skills behind AI/ML Infrastructure Engineer roles (see the career path).
Why MLOps?
As AI becomes core to modern applications, the gap between building ML models and running them in production must be bridged. MLOps applies DevOps principles — automation, versioning, CI/CD, observability — to ML-specific workflows. The good news: most of the foundation is standard DevOps, and it's all on CloudCaptain.
Stage 1: Foundations (2–3 weeks) — learn these here
- Python — the language of ML; then Python for DevOps scripting
- Docker — every model ships in a container; master the Dockerfile
- Git — plus ML-specific versioning (DVC, Git-LFS)
- Linux — GPU boxes are Linux boxes
- Basic ML concepts — training, inference, evaluation (any intro ML course works)
✅ Checkpoint: you can containerize a Python app and explain the difference between training and inference infrastructure.
Stage 2: ML Infrastructure (2–3 weeks)
- Experiment tracking & model registry — MLflow, Weights & Biases
- Data pipelines — Apache Airflow, Kubeflow Pipelines, Argo Workflows
- Feature stores — Feast, Tecton
- GPU infrastructure — cloud GPU instances, spot strategies; grounding in AWS or GCP (GCP is the most ML-native provider)
Stage 3: Model Serving & Deployment (2–3 weeks)
- Serving frameworks — TensorFlow Serving, Triton, BentoML
- Kubernetes for ML — Kubeflow, Seldon Core, KServe; know Workloads and autoscaling first
- Progressive delivery — A/B tests and canary deployments for models, using CI/CD and GitOps patterns
- Edge deployment — quantization, ONNX, model optimization
Stage 4: Production ML & LLMOps (2–3 weeks)
- Monitoring — model performance, data drift, concept drift (Evidently, WhyLabs)
- Automated retraining — trigger pipelines from drift signals
- LLMOps — serving large language models (vLLM, Ollama), prompt/version management, token-cost engineering
- FinOps for ML — GPU costs dominate; learn to measure and cut them
- Responsible AI — governance, auditability, reproducibility
Key Tools Map
| Category | Tools |
|---|---|
| Experiment Tracking | MLflow, W&B, Neptune |
| Pipelines | Kubeflow, Airflow, Argo Workflows |
| Serving | Seldon, BentoML, TF Serving, Triton |
| Monitoring | Evidently, WhyLabs, Fiddler |
| LLMOps | vLLM, Ollama, LangChain |
Prove It
- Build — deploy a model behind a REST API on Kubernetes with autoscaling, tracked in MLflow. That single project exercises every stage above.
- Interview — the ML-adjacent fundamentals get tested too: Python · Docker · Kubernetes
- Level up — pair with the Containers path for orchestration depth, or the Cloud path for GPU infrastructure economics.