AI Engineer / AI Platform

AI Engineers build the AI features people use: RAG, agents and LLM apps. AI Platform Engineers build and run what sits underneath: model serving, evals, observability and guardrails. The skills overlap so much that most teams use the titles interchangeably, and it is the biggest part of the AI job market right now. This roadmap covers both, in the order I would learn them.

Before you start

• One cloud provider to a working level (AWS, Azure, or GCP)
• Docker and Kubernetes fundamentals, including GPU scheduling basics
• Terraform for infrastructure as code
• CI/CD pipelines (GitHub Actions or GitLab CI)
• Python to a confident, practical level
• Basic observability (Prometheus, Grafana)

The roadmap

1. LLM Fundamentals

   Transformers, attention, tokens, context windows, temperature, sampling, top-k, top-p. You do not need to train models from scratch, but you need to understand how they work to operate them well.

2. Working with LLM APIs

   OpenAI, Anthropic, Google, Mistral, Cohere. Chat completions, function calling, streaming, structured outputs, rate limits, retries, fallbacks, and provider abstraction.

3. Prompt Engineering and System Prompts

   The right prompt is the difference between useless and production-ready. System prompts, few-shot examples, chain-of-thought, output formatting, prompt templating and versioning.

4. Embeddings and Vector Databases

   The retrieval layer for RAG and search. Embedding models, cosine similarity, vector indexes (HNSW, IVF), Qdrant, Pinecone, pgvector, Weaviate, Milvus.

5. RAG Systems

   The dominant pattern for grounding LLMs in your data. Chunking strategies, retrieval, reranking, citation, hybrid search, and evaluation.

6. LLM Orchestration (LangChain, LlamaIndex)

   Frameworks for building LLM applications. Chains, agents, tools, memory, callbacks. Knowing when to use a framework and when to call the API directly is part of the skill.

7. Self-Hosted LLM Serving (vLLM, TGI, Ollama)

   When you cannot use a hosted API for cost, latency, or compliance reasons. vLLM is the standard for production GPU serving with PagedAttention and continuous batching. Text Generation Inference from Hugging Face and Ollama for local development round it out.

8. LLM Observability (Langfuse, LangSmith, Phoenix)

   Logs, traces, and evaluations for LLM applications. You cannot optimise what you cannot see, and LLM apps fail in ways traditional monitoring misses.

9. Evaluations (LLM-as-Judge, RAGAS)

   The hard problem in GenAI. How do you know your LLM app is actually working? Eval frameworks, golden datasets, LLM-as-judge, RAGAS for RAG, custom metrics.

10. Guardrails and Safety

    Keeping the model within safe and intended behaviour. Input filtering, output filtering, content moderation, prompt injection protection, jailbreak prevention.

11. Agents, Function Calling, and MCP

    Where the industry is heading. Tool use, multi-step reasoning, and the Model Context Protocol for sharing tools across agents and applications.

This is the free version. The full one, with what good looks like for every skill and a production capstone project to prove it, is what I teach inside my community. Send me a DM if you want to know more.