AI engineering roadmap

This path is for building with AI: using pre-trained models through their APIs to make real things, which needs far less maths than training models from scratch. It starts with using AI as a coding partner, then follows the common 2026 order, prompting, then your own data, then agents, then making it reliable and safe, with the deep model internals kept as an optional last phase. Works in TypeScript or Python. Each phase becomes its own article.

1. 1

   Build using AI

   Read it

   Using AI as a coding partner, before you build AI into anything. The skill most people already use daily but never learn properly.

   Learning to direct a very fast junior developer, not to replace yourself with one.

   • Claude Code & agentic coding
   • CLAUDE.md / AGENTS.md
   • Plan-then-build workflow
   • MCP for dev tools
   • Reviewing AI output
   • Context engineering
2. 2

   What a language model actually is

   Read it

   Enough of the inside to stop it feeling like magic, before you build on it.

   A very well-read autocomplete that guesses the next word.

   • Tokens
   • Prediction
   • Context window
   • Temperature / top-p
   • Hallucination & knowledge cutoff
   • Open vs closed models
   • Model selection (small vs large)
   • Multimodal (vision, audio)
   • TypeScript or Python path
3. 3

   Talking to models

   Read it

   Getting useful, reliable, affordable answers out of a model through its API.

   Briefing a fast but literal assistant who only knows what you tell it.

   • LLM APIs & SDKs (Anthropic SDK, Vercel AI SDK)
   • Prompt engineering
   • System vs user prompts
   • Few-shot
   • Structured output
   • Streaming
   • Extended thinking
   • Prompt caching
   • Rate limits, retry & errors
   • Batch API
   • API key safety
4. 4

   Giving a model your own data

   Read it

   Making a model answer from your documents instead of guessing.

   Handing the assistant the exact files before you ask the question.

   • Embeddings
   • Vector databases (pgvector)
   • Chunking
   • RAG
   • Document parsing (PDF, OCR)
   • Hybrid search (BM25 + vector)
   • Reranking
   • Metadata filtering
   • Contextual retrieval
   • RAG vs long context
   • Ingestion & refresh pipeline
   • Retrieval eval (recall@k)
   • Citations
5. 5

   Making a model act

   Read it

   Letting a model use tools and take steps, safely, not just answer.

   Giving the assistant a phone, a budget, and a list of what needs your sign-off.

   • Tool use & function calling
   • Agents & agent patterns (ReAct, planner-executor)
   • Memory (short & long-term)
   • Human-in-the-loop approval
   • Permissions & sandboxing
   • MCP: using and building servers
   • Subagents / multi-agent
   • Agent SDKs & frameworks
   • Computer use & browser agents
   • Async / long-running tasks
6. 6

   AI product UX

   Read it

   The interface side: what a good AI feature feels like to use, and what a bad one gets wrong.

   The difference between a helpful shop assistant and one who never says 'I am not sure'.

   • Chat & streaming UI
   • Generative UI
   • Latency & loading UX
   • Showing citations
   • Handling 'I do not know'
   • User feedback (thumbs up/down)
   • AI SDK UI hooks
7. 7

   Making it reliable and safe

   Read it

   The part that turns a fun demo into something you can trust in front of users, and defend.

   The kitchen inspection, but for answers instead of food.

   • Evaluation & LLM-as-judge
   • Eval in CI (promptfoo, Braintrust)
   • Tracing (Langfuse, LangSmith)
   • Prompt versioning
   • Prompt injection & OWASP LLM Top 10
   • Injection scanning
   • Red teaming
   • Guardrails & content moderation
   • Output validation (Zod)
   • PII, privacy & data residency
   • Model routing & fallback
   • Semantic cache
   • Cost alerts & budgets
   • Quality monitoring
8. 8

   Beyond text

   Read it

   Voice, images, and video, once text is solid.

   Once the assistant can write, teaching it to speak, draw, and film.

   • Voice & realtime (speech-to-text, text-to-speech)
   • Image generation
   • Video generation & pipelines
   • Multimodal input
9. 9

   Going deeper (optional)

   Read it

   How models work inside, for when using them is not enough and you want to shape them.

   Learning to cook, not just to order well.

   • Neural networks
   • Transformers
   • Training
   • Fine-tuning vs prompting vs RAG
   • Local models (Ollama, llama.cpp)
   • Quantization
   • Distillation
   • RLHF (concept)
   • The maths

Articles for this path are being written. New phases and articles are added as it grows.