AI & automation · Career
From DevOps foundations to AI automation systems.
DevOps experience maps well to production AI work because AI systems still need infrastructure, observability, security, release discipline and failure handling. The transition is less about abandoning operations and more about applying those habits to a new workload class.
Why the DevOps foundation transfers
| DevOps capability | AI systems application |
|---|---|
| Kubernetes and containers | Model serving, GPU-aware scheduling, inference services and workload isolation |
| Observability | Latency, token usage, traces, evaluation signals and workflow failure visibility |
| Incident response | Tool failures, bad retrieval, degraded models, timeouts and fallback behavior |
| Security and RBAC | Data boundaries, tool authorization, auditability and tenant isolation |
| CI/CD and GitOps | Versioned prompts, evaluations, application changes and deployment controls |
| Infrastructure as code | Repeatable model, retrieval and agent environments |
Phase 1: build the foundations
Start with concepts that change how the system behaves rather than chasing every new framework. Learn enough transformer and tokenization fundamentals to reason about context windows and inference constraints. Then focus on retrieval-augmented generation, tool calling, agent state, evaluation and failure handling.
Build three small systems
- A local RAG application: ingest a controlled document set, retrieve relevant passages, answer with citations, and inspect failure cases.
- A tool-using workflow: connect a small set of explicit tools with bounded permissions, state and error handling.
- An evaluation harness: maintain test cases that run when prompts, retrieval logic, models or tool behavior change.
Build systems that can fail in ways you can observe. The debugging work is where operational understanding develops.
Phase 2: learn production patterns
Model serving
Learn the trade-offs between hosted APIs and self-hosted inference. For self-hosted work, understand serving layers such as vLLM or Ollama, resource limits, batching, autoscaling and how model size affects infrastructure choices.
Observability
Useful signals include:
- request latency and error rate
- input and output token volume
- cost or infrastructure usage per workflow
- retrieval quality and citation failures
- evaluation score changes over time
- tool-call failure and retry behaviorThe important step is connecting these signals to a user or business workflow rather than collecting telemetry without a decision attached.
Security and governance
- Keep tenant and data boundaries explicit.
- Authorize tools separately from model access.
- Log material automated actions.
- Validate external inputs and retrieved content.
- Define where human approval is mandatory.
Phase 3: build a portfolio that demonstrates operations
| Project | What it should demonstrate | Possible stack |
|---|---|---|
| Production-style RAG system | Ingestion, retrieval, generation, evaluation and monitoring | FastAPI, PostgreSQL/pgvector, an inference provider, OpenTelemetry |
| Stateful agent workflow | Tool use, state, human handoff, retries and recovery | LangGraph or another orchestrator, PostgreSQL, API service |
| AI platform service | Multi-tenancy, authentication, versioning, observability and developer experience | TypeScript or Python, OpenAPI, CLI and documentation |
A 30-day practice loop
| Week | Focus | Deliverable |
|---|---|---|
| 1 | Local inference and first retrieval workflow | A cited question-answering application over a controlled document set |
| 2 | Tool use and workflow state | A small automation with explicit failure and handoff paths |
| 3 | Evaluation and CI | Regression tests that run whenever workflow behavior changes |
| 4 | Deployment and observability | A deployed service with logs, metrics, traces and a short operating runbook |
Position the transition around systems, not hype
A stronger professional narrative is that you build and operate AI-enabled systems: you understand the application workflow, the infrastructure behind it, the evidence needed to evaluate changes, and the controls required when automation touches production work.
Use the structured playbook
The AI Automation Career Playbook turns this roadmap into a longer implementation-focused guide for operators moving into AI automation work.
Explore the playbook