“AI engineering” — building products on top of language models — is now a core skill for software developers. This path assumes you can already program and focuses on what is new.

1. Call a model from code

Read the official documentation of one or two providers and make your first API calls. Learn about tokens, context windows, temperature, system prompts and rate limits, and how pricing works per token.

2. Structured output

Real applications need data, not prose. Practice asking for JSON that matches a schema, and validate it in code. Most providers now offer structured-output or schema features.

3. Retrieval-augmented generation (RAG)

Learn to split documents into chunks, create embeddings, store them in a vector index and retrieve the most relevant passages for each question. See our RAG guides.

4. Tool use and agents

Let the model call functions: search, databases, calculators, your own APIs. Then study agent loops, where the model plans, acts and observes results repeatedly — and the safeguards they need.

5. Evaluation

The skill that separates prototypes from products. Build a set of test cases, define what “good” means, and measure every change to prompts, models or retrieval. See our Evals & Observability coverage.

6. Open models and fine-tuning

The Hugging Face LLM Course teaches how transformer models work in code, how to run open-weight models and how to fine-tune them.

7. Production concerns

Latency, caching, cost control, monitoring, prompt-injection defenses and data privacy. Ship something small, watch real usage and iterate.

Suggested project ladder

  1. A command-line Q&A tool over your notes (RAG).
  2. A support assistant that can look up orders (tool use).
  3. An evaluation harness that compares two models on your own test set.

Common pitfalls for developers

  • Skipping evaluation: without a test set, you can’t tell whether a prompt change helps or hurts.
  • Trusting model output as data: always validate structured outputs and handle failures.
  • Ignoring cost and latency until production, when they become painful.
  • Over-engineering agents when a simple pipeline would be more reliable.
  • Neglecting security: prompt injection, data leakage and over-permissive tools are real risks — see our prompt injection explainer.

Model providers publish detailed documentation and cookbooks with examples for each step on this path. Our guides cover your first API call, structured outputs and function calling, retrieval-augmented generation and building an evaluation set.

How long it takes

A developer comfortable with an existing programming language can build a working prototype within days, reach a solid understanding of RAG and tools within a few weeks, and gain production-level skills — evaluation, monitoring, security and cost control — over a few months of real projects.

A first project in one weekend

Build a small “ask my documents” app:

  1. Saturday morning: call a model API from a script and get structured JSON output.
  2. Saturday afternoon: load a few documents, split them into chunks, create embeddings and store them in a simple vector index.
  3. Sunday morning: retrieve relevant chunks for a question and pass them to the model with instructions to cite sources.
  4. Sunday afternoon: write 15 test questions with expected answers, run them, and improve chunking or prompts based on failures.

You’ll touch most of the skills on this path — and end with something you can show.

Related: the free learning path to understand AI for broader foundations.

What to watch next

  • Agent frameworks and standards such as the Model Context Protocol.
  • Better evaluation and observability tools for LLM applications.
  • Falling costs that make new applications viable.
  • Security guidance for AI agents with tool access.

Key takeaways

  • Start by calling a model from code, then add structured output, retrieval and tools.
  • Evaluation is essential from the beginning, not an afterthought.
  • Consider open models and fine-tuning when you need control or specialization.
  • Production requires attention to cost, latency, monitoring and security.
  • Build a ladder of increasingly complex projects.

Frequently asked questions

Which programming language is best for building with LLMs?

Python and TypeScript have the richest ecosystems, but any language that can make HTTP requests works.

Do I need to know machine learning?

Not to start. Understanding model behavior, evaluation and limitations matters more than training models.

Should I use a framework?

Frameworks can speed development, but understanding the underlying API calls first makes debugging easier.