Free course
Build a real AI app from scratch.
ChatGPT shipped in late 2022. By mid-2023 every product had an "AI" sticker on it, and most of the engineers who built them learned on the fly. If you haven't started yet, you didn't miss the boat.
6 lessons in Chapter 1 · Python & TypeScript · <$0.10 in API credit
You're not late.
The first wave of AI builders learned by Googling, copy-pasting from random Twitter threads, and burning thousands of dollars in OpenAI credit figuring things out. The second wave doesn't have to.
Every concept here costs less than a cent of API credit to run. No GPU. No fine-tuning. No ML degree. If you can read Python or TypeScript and you know what an HTTP request is, you can ship.
What this course is.
A free, open-source course that takes you from your firstchat.completionscall to a real production AI app. The patterns are vendor-neutral. Examples use OpenAI and Respan as a concrete reference; the same shape applies to any provider.
- 6 runnable examples: LLM call, metrics, prompt roles, providers, tool calls, structured output.
- Two languages. Every example in Python and TypeScript, side by side.
- Cents to run. Default model is
gpt-4o-mini. A typical lesson costs under $0.01. - Real production layers. By Chapter 1.6 you have a working agent with a gateway, prompt registry, traces, and an eval suite.
60-second glossary.
The terms you'll keep seeing.
- LLM
- Large language model. Trained on huge amounts of text. Takes text in, produces text out. GPT-4o, Claude Sonnet 4.6, Gemini 2.5 are LLMs.
- Prompt
- The input you send. Usually a system message (instructions) and a user message (the question).
- Token
- A chunk of text. ~0.75 words on average. Providers charge per input token and per output token.
- Context window
- The maximum tokens a model can read in one call. Exceed it and the API errors out.
- Hallucination
- When the model confidently says something that is not true.
- RAG
- Retrieval-Augmented Generation. Fetch your own documents first, then put them in the prompt so the model answers from real sources.
- Tool call
- When the model asks your code to run a function (read a database, call an API) and waits for the result.
- Agent
- An AI system that decides its own next step in a loop. A workflow is the simpler cousin where the steps are fixed.
- Trace / span
- A record of what the system actually did. A trace covers one request; spans are the steps inside it.
- Eval
- An automated way to score the quality of an LLM output. Without one, you have no idea if a prompt change helped or hurt.
Start here.
Each lesson builds on the last. By the end of Chapter 1 you have a working production AI app.
Chapter 1
- 1.1Your First LLM Call->
What an LLM API call actually does, what each piece of a prompt is, what tokens cost, and the six things that break the moment you try to run a real product on top of one call.
- 1.2What Is an LLM Gateway?->
An LLM gateway is a single URL between your app and every LLM provider. The pattern gives you logging, fallbacks, caching, multi-provider routing, and per-customer cost caps without rewriting your application.
- 1.3Prompt Management for LLM Apps->
Prompts are artifacts, not strings in source code. Versions, environments, A/B testing, and one-click rollback. How to write a system prompt, add variables, and deploy a prompt registry safely.
- 1.4LLM Tracing and Workflows->
When one LLM call is not enough, you compose multiple calls into a workflow. Without LLM tracing, debugging is impossible. Tracing is how you see what your AI actually did.
- 1.5LLM Evals: Measuring AI Quality->
Without LLM evals, you change a prompt and have no idea whether it made things better. With evals, you have a number per change and a CI gate that blocks regressions.
- 1.6AI Agents and Tool Use->
Single call, workflow, agent. Three LLM architectures with different trade-offs. When the next step has to be decided by the model rather than your code, you have a real AI agent. Otherwise you have a workflow pretending to be one.