AI Engineering Roadmap for Experienced Developers
A structured, hands-on path from full-stack development to production AI engineering: 4 areas, 11 modules, 25 vetted resources, roughly 57 hours of study and implementation. Free, and updated as I actually work through it (last updated November 21, 2026).
How language models behave and how to talk to them, before building anything on top of them.
What a fictional learner needs before anything else: tokens, attention and why sampling settings change an answer.
A fictional transformer explainer (opens in a new tab), Completed
Video, estimated 1hFictional notes on tokenization, Completed
Text, estimated 30mA sampling playground, Skipped
Repo
A fictional course on prompt patterns (opens in a new tab), Completed
Link, estimated 4hStructured output drills, Skipped
Repo, estimated 1.5h
Turning text into vectors and reading the space they live in.
- Distance and similarity
- Where embeddings mislead
A fictional talk on embedding geometry (opens in a new tab), In progress
Video, estimated 45mA fictional talk about how meaning clusters in vector space. Notes so far cover cosine similarity and normalisation.
A vector maths refresher (opens in a new tab), Planned
TextReading the fictional vector search handbook, Completed
Finding the right passages for a model to answer from, and measuring whether it did.
Keyword and vector search side by side, before any retrieval pipeline is built on top of them.
A fictional BM25 walkthrough (opens in a new tab), Completed
Link, estimated 50mA fictional playlist on hybrid search (opens in a new tab), In progress
Playlist, estimated 3hA fictional playlist that combines BM25 with dense vectors.
- Score fusion
- Weighting the two signals
Fictional lecture notes on ranking (opens in a new tab), Paused
Drive, estimated 2h
Chunking experiments, Planned
Repo, estimated 5h
A fictional guide to retrieval metrics (opens in a new tab), Planned
Text, estimated 1hA fictional workshop on golden sets (opens in a new tab), Planned
Video, estimated 2hFaithfulness checks, Planned
RepoA fictional paper on grading answers (opens in a new tab), Planned
Link, estimated 45mReading the fictional evaluation field guide, In progress
Nothing planned here yet.
No resources yet.
Tools, loops and guardrails for agents that run unattended.
A fictional tutorial on tool schemas (opens in a new tab), Completed
Link, estimated 40mHow a fictional assistant describes its tools so a model can call them with valid arguments.
A function router, Completed
Repo, estimated 8h
Planning, acting and checking in a loop, and knowing when to stop.
A fictional video on planning loops (opens in a new tab), Planned
Video, estimated 35mA state-machine agent, Planned
Repo, estimated 10hA fictional book on reflection patterns, Planned
BookOn the bookshelf
A fictional guardrail checklist (opens in a new tab), Skipped
Text, estimated 20mA red-team exercise, Planned
Repo, estimated 4hA planned exercise: try to break the fictional guardrails before anyone else does.
Nothing planned here yet.
No resources yet.
A placeholder for papers to follow once the practical areas are done.
No modules yet.
Learning approach
The roadmap follows three principles:
- Build on existing software-engineering experience instead of starting again as a junior developer.
- Prioritize production systems, measurable quality, and reliability over framework knowledge or isolated demos.
- Spend approximately 40% of learning time on study and 60% on implementation, experimentation, and documentation.
Some resources are long-term commitments (like books) and topics often overlap across resources. You'll encounter concepts you've already learned, but also discover new perspectives and extensions. This means you'll spend less time on each resource than if you studied them in complete isolation - the knowledge compounds.
The dots beside each area show its recommended share of the total learning effort.
Portfolio Project Ideas
Additional project ideas to strengthen your portfolio beyond the core roadmap.
Why it matters
A small retrieval project teaches chunking, ranking and grounding on text you know well, so wrong answers are easy to spot.
Key topics
- Retrieval
- Evaluation
How to start
- Export a handful of fictional notes as Markdown.
- Split them into passages and index them with keyword search first.
- Add vectors only where keyword search misses.
Why it matters
Two tools are enough to practise tool selection, argument validation and recovering from a failed call.
Key topics
- Agents
- Tool use
- Guardrails
How to start
Start with a read-only
list_eventstool and addcreate_eventbehind a confirmation step.Why it matters
Twenty labelled examples and a pass threshold turn prompt changes from guesswork into measurements.
Key topics
- Evaluation
Stay Current
Resources to keep up with the rapidly evolving AI landscape. The field moves fast — these help you stay informed without getting overwhelmed.
Example AI Weekly (opens in a new tab)
newsletterA fictional weekly newsletter with one paper, one tool and one practical write-up.
Conversations about measuring language-model systems in production.
A local study group
communityA fictional monthly meeting without a website.
Frequently Asked Questions
Common questions about the roadmap, its scope and who it's for.
A structured, public learning roadmap documenting the transition from experienced full-stack development to production AI engineering. It covers programming fundamentals, LLM and ML theory, RAG, AI agents, LLMOps, evals and safety, platform architecture and AWS AI certification, organized into sequential areas.
Developers who already have several years of software engineering experience (TypeScript, backend frameworks, system design) and want to extend that expertise into AI-native development, rather than starting from scratch as a beginner programmer.
Each area is scoped around building something real — a RAG service, an agent runtime with tool orchestration, an evaluation harness with CI regression checks — instead of isolated tutorials. Time is split roughly 40% study and 60% implementation and documentation.
A personal learning log kept public and structured like a course. The roadmap, resources and diary reflect real day-by-day progress, including what worked and what didn't, rather than a polished curriculum sold to students.
The roadmap progresses toward a production-oriented RAG service, a multi-tool agent runtime with audit trails and human-approval gates, an LLMOps dashboard with cost and quality monitoring, and a small shared AI platform supporting multiple agents — each built as a real project, not a toy demo.
Python and FastAPI for AI-native backends, pgvector for retrieval, LangGraph and CrewAI for agent orchestration, structured-output and tool-calling patterns across LLM providers, plus observability, evaluation and safety tooling — building on an existing TypeScript and Next.js background.
Additional Terms to Explore
Terms from the AI world worth knowing about, but not directly part of this roadmap.
- Context window
The amount of text, measured in tokens, that a model reads in one request: the prompt, any retrieved passages and the answer so far.
- Embedding
A list of numbers that represents the meaning of a text, so that similar texts lie close together. Used for semantic search.
- Grounding
Answering only from supplied sources and pointing to them, instead of from what the model happens to remember.
- Eval
A repeatable test of a model-based system: fixed inputs, an expected behaviour and a pass threshold.