Roadmap · Zero to Production
The Enterprise AI Roadmap
Not sure where to start? This is the exact sequence I'd follow today — what to learn, what to read, and which project to build at every stage. No skipping: each stage's project is the foundation of the next.
One rule before you begin: building beats watching. Every stage ends with a project because the projects are the learning — and by stage 6 you'll have a portfolio that demonstrates production-grade skills, not tutorial completion.
Stage 01 — Foundations
LLM Fundamentals & Prompting
Week 1–2
Learn
How LLMs actually work — tokens, context windows, temperature, why models hallucinate. Structured prompting: roles, constraints, few-shot examples, structured outputs (JSON mode). You cannot architect what you don't understand at this level.
Read / Watch
Subscribe to @aiarchitectpbr — new videos Tue & Fri. Start with AI Agents vs Agentic AI vs Workflows to get the vocabulary right from day one.
Build
A CLI assistant that takes a messy meeting transcript and returns structured JSON — decisions, owners, deadlines. Forces you to master structured outputs, the foundation of everything that follows.
Stage 02 — Retrieval
Production RAG
Week 3–4
Learn
Chunking strategies, embeddings, vector databases, hybrid search (BM25 + vectors with RRF), cross-encoder reranking, and evaluation with golden sets. This is the most in-demand enterprise AI skill, and the most commonly botched.
Read / Watch
Why 95% of RAG Demos Fail in Production — the seven failure modes — then RAG vs Fine-Tuning vs Long Context for when retrieval is even the right tool.
Build
A document Q&A system over 50+ real PDFs (annual reports work well) with hybrid search and reranking. Then build a 50-question golden set and measure your accuracy — measuring is the skill.
Stage 03 — Agents
Single AI Agents
Week 5–6
Learn
Agent anatomy: the reason–act–observe loop, tool calling, ReAct, guardrails, termination conditions. When an agent is justified versus a workflow — autonomy is a cost you spend only where judgment is needed.
Read / Watch
AI Agents vs Agentic AI vs Workflows for the decision framework, then Prompt Injection: Your Agent's Biggest Production Risk before you give any agent real tools.
Build
A research agent that takes a company name, calls 3+ tools (search, your Stage-2 RAG system, a calculator), and produces a sourced one-page brief — with a hard step limit and an "I don't know" path.
Stage 04 — Orchestration
Multi-Agent Systems · MCP & A2A
Week 7–9
Learn
Supervisor patterns, LangGraph state machines, parallel fan-out, checkpointing, failure recovery. Then the protocol layer: MCP servers for tool access, A2A for agent-to-agent delegation — the 2026 enterprise foundations.
Read / Watch
Multi-Agent Orchestration Patterns That Survive Production and MCP vs A2A: Which Protocol Should Your Enterprise Bet On?, plus Agent Memory for state that persists.
Build
A supervisor + two specialists (researcher, writer) orchestrated in LangGraph with checkpointed state — and expose one of your tools as an MCP server so any LLM can use it.
Stage 05 — Specialisation
Fine-Tuning & Azure AI Foundry
Week 10–11
Learn
LoRA/QLoRA: when fine-tuning beats prompting (behaviour problems, not knowledge problems), dataset preparation, held-out evaluation, DPO alignment. Plus deploying it all on enterprise infrastructure with governance.
Read / Watch
LoRA Fine-Tuning for Finance: +31 Points — the data-first method — and Azure AI Foundry: The Honest Enterprise Review for the deployment reality.
Build
LoRA fine-tune a 7B model on a classification task with 500+ labelled examples. Score it against the prompted base model on a held-out set — if you can't show the delta, you haven't finished.
Stage 06 — Production
LLMOps, Evals & Cost
Week 12+
Learn
Tracing, token cost attribution, evaluation gates in CI, observability, incident response. The discipline that separates engineers who demo from architects who ship — and the skill enterprises actually pay for.
Build
Take your Stage-4 system to production grade: full tracing, a cost-per-task dashboard, an eval suite wired into CI that blocks regressions. This is your portfolio capstone — the artifact that gets you hired or promoted.
Want this roadmap, taught live?
My 8-week Agentic AI course covers stages 3–6 with live sessions, code reviews and real enterprise projects — batch 01 starts 7 July, 50 seats.
View the course →No spam. Unsubscribe anytime. New Tuesdays.