Python, APIs and LLM basics
You cannot run what you do not understand. Before you operate an AI system you need to read and write Python, know how a call to a model provider works over HTTP, and understand how a language model behaves once tokens, cost and latency come into it. Then you build the workload itself: a small service and a minimal agent, so the thing you put behind a gateway, a server and a monitor is one you actually understand, not a black box someone handed you.
Python for AI
APIs for AI
LLM Fundamentals
OpenAI Agents SDKModel access and gateways
The first thing a platform needs is one place every service calls the model through, instead of each app talking to each provider directly. This stage builds that gateway: a single endpoint with retries, fallbacks, caching, budgets and cost tracking, so you can change model, add a provider or cap spend without touching a single application.
LiteLLMServing and scale
A hosted API is not the only way to run a model, and it is rarely the cheapest at scale or the option a customer with private data will accept. This stage is about serving models yourself and running the agent behind an endpoint that can take load: an inference server, a real deployment, and the orchestration that keeps long-running work alive.
vLLM
LangGraphObservability and cost
Once traffic is real, the questions change from does it work to why did that run cost so much and where did the latency go. This stage is about seeing inside a live system: tracing every model and tool call, linking a result back to the exact prompt and version, and watching cost and latency closely enough to act before a bill or a user complains.
LangfuseReliability and safety in production
A platform is judged on the bad day, not the good one. This stage adds the layers that keep a live system safe and predictable: checks on what the model returns, a human in the loop before anything irreversible, a sandbox for any code it writes, one place that allows or denies every tool call, and a clear map of where user data travels so it never leaks into a log or a trace.
LangGraph
Guardrails AI
E2B
Agent Governance ToolkitDeploy and hand over
The last stage puts it all together into something a team can run without you: the agent served, a UI in front of it, traces and budgets in place, and a handover that lets the customer's own engineers operate and change the system after you go.
Vercel AI SDKLangGraphPlatform engineer
6 stages, 19 frameworks and 2 projects. Finish them and you will have shipped the kind of systems these teams hire for, with the lessons to back every one.
Browse all frameworksQuestions about this roadmap
What does an AI platform or LLMOps engineer do?
They run the infrastructure that AI features depend on: a gateway in front of every model provider, serving and scaling, observability and tracing, cost and rate limits, guardrails, and deployment. It is MLOps discipline applied to language models, where cost, latency and safety matter as much as accuracy.
Can I start this roadmap as a beginner?
Yes. The first stage teaches Python from your first line of code, then APIs and how language models work. Every later stage assumes only the stages before it.
Do I need an API key or a paid model account?
No. The courses use small stand-in models you write yourself, or a small open model that runs on your own computer, so no lesson needs an API key or a paid account. The gateway and serving stages are where you learn how a real hosted model would slot in.
Some stops say coming soon. Why?
A few frameworks on the serving and orchestration stages do not have a full course yet, so those stops link to the official docs and are marked done by hand. The rest of the roadmap is live lessons you can take today, and the coming-soon courses drop into the same stops as they are written.
How is this different from the AI Forward Deployed Engineer roadmap?
That roadmap builds an AI system and ships it to a customer. This one is about the platform underneath: it goes deep on gateways, serving, observability, cost and reliability, which the forward deployed roadmap only touches when it deploys.