AI Engineer Roadmap 2026
The complete journey from software engineering foundations through production AI systems. Master Python, APIs, LLM fundamentals, prompt engineering, RAG, tool calling, agents, evaluation and production deployment.
Visual Roadmap
10 stages · click any stage to jump
Foundation
Data, databases, containers and configur…
Python & APIs
Build robust backend services that call,…
LLM Fundamentals
Tokens, context windows, sampling, prici…
Prompt Engineering
Instructions, examples, structured outpu…
RAG
Ingestion, chunking, embeddings, retriev…
Tool Calling
Let models call your functions safely.
Agents
Agent loops, memory, planning and knowin…
Evaluation
Measure quality before and after every c…
Production AI
Observability, cost, reliability, securi…
Capstone
Ship a production AI system end to end, …
Stage by stage
0/22 chapters read · 0%1Foundation2 weeks0/2 readData, databases, containers and configuration: the plumbing every AI system sits on.
You will be able to
- Model data in SQL and choose between relational, document and vector storage
- Containerise a service and run it with its dependencies in one command
- Keep API keys and configuration out of code with environment variables and a secrets store
- Use Git branches and pull requests for every change
Ready to move on when: You can clone a repository, run a Python service and its database with one command, and nothing secret is committed.
2Python & APIs1 week0/1 readBuild robust backend services that call, and are called by, other systems.
You will be able to
- Write async Python that calls HTTP APIs with timeouts
- Validate inputs and outputs with typed models
- Retry transient failures with exponential backoff, and fail fast on the rest
- Test an API with unit and integration tests
Ready to move on when: You have a typed HTTP API with input validation, timeouts, retries and tests that run in CI.
3LLM Fundamentals1 week0/2 readTokens, context windows, sampling, pricing and choosing between hosted and open models.
You will be able to
- Explain tokens, context windows and why outputs vary between runs
- Estimate the cost and latency of a request before you send it
- Choose between a hosted API and an open model you run yourself
Ready to move on when: You can estimate a feature's monthly model cost and explain why the same prompt gives different answers.
4Prompt Engineering1 week0/2 readInstructions, examples, structured outputs and streaming.
You will be able to
- Write prompts with clear instructions, examples and output formats
- Get schema-valid JSON every time, and handle the times you do not
- Stream responses to users without breaking partial structured output
- Version prompts and test them against a fixed set of inputs
Ready to move on when: Your prompts are versioned, return schema-valid output on a 20-input test set, and stream to the UI.
5RAG3 weeks0/5 readIngestion, chunking, embeddings, retrieval and reranking, with citations.
You will be able to
- Decide when RAG is the right tool, and when it is not
- Ingest messy documents and chunk them so answers stay intact
- Embed, index and retrieve with vector and hybrid search
- Rerank results and answer with citations to the source
- Measure retrieval recall separately from answer quality
Ready to move on when: Your assistant answers with citations, and you can show retrieval recall on a labelled question set before and after reranking.
6Tool Calling1 week0/1 readLet models call your functions safely.
You will be able to
- Define tools with precise schemas and descriptions
- Validate tool arguments before running anything
- Return tool errors the model can recover from
Full series: Building AI AgentsPractise
Ready to move on when: The model calls your tools with validated arguments, and recovers sensibly when a tool fails or times out.
7Agents2 weeks0/2 readAgent loops, memory, planning and knowing when to stop.
You will be able to
- Build a reason-act-observe loop from scratch, without a framework
- Give agents short-term and long-term memory
- Set stopping conditions, step budgets and cost limits
- Trace every step an agent takes
Full series: Building AI AgentsPractise
Ready to move on when: Your agent completes a multi-step task within a step budget, stops cleanly when stuck, and leaves a trace of every action.
8Evaluation2 weeks0/3 readMeasure quality before and after every change.
You will be able to
- Build a labelled evaluation set from real or realistic inputs
- Score outputs with code checks, and with LLM judges you have calibrated
- Run evaluations in CI and catch regressions before release
- Monitor quality in production
Full series: LLM EvaluationPractise
Ready to move on when: Changing the model or a prompt runs your evaluation suite automatically and blocks the merge if quality drops.
9Production AI2 weeks0/4 readObservability, cost, reliability, security and context engineering.
You will be able to
- Trace requests end to end, with cost, latency and quality per request
- Cut costs with caching, routing and smaller models where they are good enough
- Survive provider outages with timeouts, fallbacks and circuit breakers
- Defend against prompt injection and data leaks
Ready to move on when: You can show cost per request, latency percentiles and error rates for your app, with a fallback model and prompt-injection defences in place.
10Capstone1 weekShip a production AI system end to end, and prove it works.
You will be able to
- Scope a project small enough to finish and big enough to matter
- Ship with evaluations, observability and a written trade-off log
- Present results with evidence, not demos alone
Practise
Ready to move on when: Your capstone meets the brief below, and someone else can run it from the README.
Capstone: Ship a cited support assistant
Build an assistant over a real documentation set, such as an open-source project's docs, that answers questions with citations and can open a support ticket through a tool. Deploy it, measure it and write up what broke.
What to deliver
- A repository with a README that runs the system in one command, and an architecture diagram
- An evaluation set of at least 50 questions with expected sources, and the scores
- A dashboard or report of cost per request, latency percentiles and error rate
- A 3-minute demo video, including a question it should refuse
- A short write-up of three failure modes you found and how you fixed them
What strong work looks like
- Correctness
- Answers cite the right source; evaluation scores are reported honestly, failures included.
- Reliability
- Timeouts, retries and a fallback model are in place, and the tool call handles errors.
- Cost and observability
- Every request is traced, and cost per answer is known and justified.
- Safety
- Tested against prompt injection in retrieved documents, with the result recorded.
- Communication
- A reviewer understands the design and the trade-offs from the README in five minutes.
Start from: RAG Assistant (project) · RAG System (pattern)
What you'll be able to do
Series in this roadmap
AI Engineer roadmap FAQ
How long does it take to become a AI Engineer?
About 16 weeks of focused study on this roadmap, which has 10 stages. Software developers entering AI engineering who want a structured path to shipping production AI systems.
What does the AI Engineer roadmap cover?
In order: Foundation, Python & APIs, LLM Fundamentals, Prompt Engineering, RAG, Tool Calling, Agents, Evaluation, Production AI, Capstone.
What will I be able to do after the AI Engineer roadmap?
Build and deploy production LLM applications. Implement RAG systems with retrieval, reranking and evaluation. Design and ship AI agents with tool calling. Apply production patterns: observability, caching, fallbacks, guardrails. Diagnose and debug failing AI systems.
How should I split my time on the AI Engineer roadmap?
At 8 to 10 hours a week: Foundation 2 weeks, Python & APIs 1 week, LLM Fundamentals 1 week, Prompt Engineering 1 week, RAG 3 weeks, Tool Calling 1 week, Agents 2 weeks, Evaluation 2 weeks, Production AI 2 weeks, Capstone 1 week. Move on when you meet each stage's "ready to move on" checkpoint, not when the time is up.
What should I build to prove I am ready for AI Engineer roles?
Ship a cited support assistant: Build an assistant over a real documentation set, such as an open-source project's docs, that answers questions with citations and can open a support ticket through a tool. Deploy it, measure it and write up what broke. Deliver 5 things, including a repository with a README that runs the system in one command, and an architecture diagram and an evaluation set of at least 50 questions with expected sources, and the scores.
Is the AI Engineer roadmap free?
Yes. Every roadmap, series and progress tracker on Learn AI Engineering is free, and progress is saved in your browser without an account.