Skip to content

AI Engineer Roadmap

In one line: A phase-by-phase plan (roughly 6 to 18 months part-time, depending on where you start) from programming foundations to shipping production LLM systems, with what to learn, what to build, and which certifications are worth it at each stage. Last reviewed: 2026-10-21

The short version

  • An AI engineer builds products on foundation models: prompts, retrieval, tools and agents, evaluation, and production operations. The role is defined in the AI careers map.
  • The roadmap has six phases: foundations, LLM fundamentals, core building blocks, production, depth, specialisation. Each phase ends with a project you can show.
  • If you already work as a software engineer, you can skip most of phase 0 and reach phase 3 in a few months of evenings.
  • Projects beat certificates for this role. Certifications help in phases 3 to 5 if you work with a specific cloud or vendor, or want to move into solutions or consulting work.
  • Re-check the certification picks before you book: several popular AI exams were retired in 2026 (noted below).

The phases at a glance

flowchart LR
    P0[0. Foundations: Python, Git, APIs, SQL] --> P1[1. LLM fundamentals]
    P1 --> P2[2. Core building blocks: RAG, tools, structured output]
    P2 --> P3[3. Production: evals, observability, cost, security]
    P3 --> P4[4. Depth: agents, fine-tuning, serving]
    P4 --> P5[5. Specialise]
Phase Time (part-time) You can show
0. Foundations 0-3 months A small Python web API on GitHub
1. LLM fundamentals 2-4 weeks A chatbot with streaming and a cost meter
2. Core building blocks 1-2 months A RAG app and an extraction pipeline
3. Production 1-2 months An eval harness in CI, traces, a red-team report
4. Depth 2-3 months An agent with MCP tools; a LoRA fine-tune; a self-hosted model
5. Specialise Ongoing A substantial project in your chosen niche

Phase 0: foundations (0-3 months, skip what you know)

Learn - Python: functions, classes, packages, virtual environments, async basics. See Python for AI. - Git and GitHub; the command line. - HTTP APIs and JSON; build and call a REST API (FastAPI is a good choice). - SQL basics. - Enough maths to read diagrams and papers: vectors, dot products, probability basics. You do not need calculus to start building. - Cloud basics on one provider.

Optional certification: a cloud fundamentals exam if you are new to cloud, for example AWS Cloud Practitioner or Google Cloud Digital Leader.

Build: a small web API that stores and returns data, deployed somewhere public.

Phase 1: LLM fundamentals (2-4 weeks)

Learn - What an LLM is and is not: LLM basics, tokenization, sampling and decoding, context windows. - How transformers work at a high level: transformer architecture and the original Attention Is All You Need. - How chat models are trained: InstructGPT / RLHF, and reasoning models. - Calling APIs: first LLM API call, API keys and env vars. - Prompt engineering and hallucinations. - Running a model locally: run a model locally.

Optional certification: NVIDIA Generative AI LLMs Associate (NCA-GENL) or AWS AI Practitioner (AIF-C01) to structure your study. Both are foundational.

Build: chatbot in an afternoon.

Phase 2: core building blocks (1-2 months)

Learn - Embeddings and vector search and RAG; the RAG paper. - Structured outputs. - Tool use and function calling; the ReAct and Toolformer papers. - Prompt caching and context window management. - Fine-tuning vs RAG: when each applies. - Multimodal models: images and PDFs as input.

Build: build a RAG pipeline and structured extraction pipeline.

Phase 3: production (1-2 months)

This phase is what separates demo builders from people companies trust with production systems.

Learn - Evals for LLMs and LLM-as-a-judge. - LLM app architecture: gateways, fallbacks, guardrails. - LLMOps and observability: tracing, metrics, release practices. - Cost and latency control. - Security: OWASP LLM Top 10, prompt injection explained, guardrails and safety. - Responsible AI: responsible AI basics.

Build: set up an eval harness wired into CI for your phase 2 project, and the prompt injection red team against it. Write up the results.

Certifications worth considering here (pick by your stack): - Claude: Claude Certified Developer - Foundations (CCDV-F) or Claude Certified Architect - Foundations (CCAR-F). - Databricks: Generative AI Engineer Associate. - Oracle: OCI Generative AI Professional.

Phase 4: depth (2-3 months)

Learn - Agents explained, agentic loops, agent architectures, MCP, agent security. - Advanced RAG patterns. - Fine-tuning: LoRA, QLoRA, DPO. - Serving: inference servers, quantization and distillation, GPUs for AI, PagedAttention / vLLM.

Build: build a Claude agent with MCP, fine-tune with LoRA, run Llama on a single GPU.

Certifications worth considering here: - AWS: Generative AI Developer - Professional (AIP-C01). - NVIDIA: Agentic AI Professional (NCP-AAI) or Generative AI LLMs Professional (NCP-GENL). - Google Cloud: Professional Machine Learning Engineer. - Claude: Claude Certified Architect - Professional (CCAR-P).

Phase 5: specialise (ongoing)

Pick one direction and go deep:

Direction Focus Next reading
Agents and automation Long-running agents, computer use, evaluation of agents Agent benchmarks, METR time horizons
Search and knowledge Retrieval quality, GraphRAG, enterprise search GraphRAG, dense retrieval
Model customisation Fine-tuning, distillation, post-training Tulu 3, knowledge distillation
Inference and infrastructure Serving at scale, GPU efficiency LLM serving, inference economics
AI security Red teaming, agent security LLM red teaming, AI threat modeling
Voice and multimodal Realtime voice, vision, documents Voice agent, multimodal models

Certification notes (as of 2026-10)

Several certifications that older roadmaps recommend are no longer available:

  • AWS Machine Learning - Specialty (MLS-C01): retired April 2025.
  • AWS Machine Learning Engineer - Associate (MLA-C01): retired 28 September 2026; a successor, MLA-C02, was in beta with no general availability date announced.
  • Microsoft Azure AI Fundamentals (AI-900) and Azure AI Engineer Associate (AI-102): retired 30 June 2026. Azure Data Scientist Associate (DP-100): retired 1 June 2026.

For everything still available, prices and levels, see the AI certifications guide.

Career progression

Stage Typical scope
Junior / early career Builds features with guidance; owns prompts and tests for one feature
Mid-level Owns an AI feature end to end: design, evals, launch, monitoring
Senior Designs systems across features; sets eval and release standards; mentors
Staff / principal Sets AI architecture and platform direction across an organisation

Pay varies widely by country, company and level; use current salary-survey data for your market rather than any single figure.

How to show your work

For each project: a public repository, a README with the problem, architecture diagram, an eval table (what you measured, the numbers, what failed), cost per task, and what you would do next. Link them from your CV. Interviewers will ask about trade-offs; see AI engineer interview and AI system design interview.

Common mistakes

  • Tutorial loops. After phase 1, every week should produce something you built.
  • Skipping evals. It is the skill that most separates candidates.
  • Framework-first learning. Learn the raw API loop before adopting a framework, so you can debug it.
  • Collecting every certificate. One or two aligned with your stack is plenty.
  • Ignoring security. Agents with tools are attack surfaces; employers know it.

Next steps

Sources

  • Zero to Hero, AI/ML Engineer Roadmap (phases, role definition, skills; used as raw material): ai-engineer-roadmap.md
  • AWS, Machine Learning Engineer - Associate (MLA-C01 retirement, MLA-C02 beta): https://aws.amazon.com/certification/certified-machine-learning-engineer-associate/
  • Microsoft Learn, retired certification exams (AI-900, AI-102, DP-100): https://learn.microsoft.com/en-us/credentials/support/retired-certification-exams
  • Anthropic certification program (via the Zero to Hero Anthropic index; Pearson VUE delivery, launched March 2026): ../../certifications/anthropic/README.md