The AI Engineer Path

Scrimba's AI engineer path, interactive in the browser editor throughout. It leads from the first API requests to a language model through streaming UIs, prompt and context engineering and a hand-built RAG pipeline (embeddings, chunking, vector database in Supabase) to agents: first built from scratch following the ReAct pattern, then via OpenAI function calling and finally with the Vercel AI SDK including tool calling, structured outputs and agentic retrieval routing. On top of that: open-source models (Hugging Face, Transformers.js in the browser, Ollama locally), deployment with staging, a health endpoint and alerting, a custom MCP server, and multimodality with image generation and vision. Several standalone solo projects are included — PollyGlot, PopChoice and an AI travel agent.

Contents & competencies

• Model selection and setup: the AI marketplace and recommended models, your own AI client, building the first API requests and then rebuilding them from scratch
• Costs and limits under control: tokens, token costs, context windows, model snapshots and the messages array
• Prompting in practice: system message, few-shot prompting, temperature and top P, JSON and structured outputs, the web search tool
• Streaming UIs: streaming responses, rendering markdown safely, handling errors and carrying the context of an ongoing conversation
• Context engineering: building system prompts deliberately, managing the context window and summarising context when it grows too large
• RAG from the ground up: generating embeddings, chunking documents, storing them in a Supabase vector database and retrieving them by similarity search — including multiple matches and error handling
• Building agents yourself: ReAct prompting with your own loop logic and response parsing, then the same agent via OpenAI function calling with automatic tool calls
• Vercel AI SDK: basics, structured outputs, tool calling and agentic retrieval routing — each with its own challenge — through to a complete customer support agent with web search
• Open-source models: Hugging Face inference for summarisation, classification and image editing, models in the browser with Transformers.js, running locally with Ollama
• Model Context Protocol: architecture and core concepts, your own MCP server with tools and resources, transports and MCP clients
• Multimodality: image generation including prompting, response formats, size/quality/style and image editing, plus vision analysis with GPT-4
• Deployment: Node environment, build command on Render, smoke test, domains, staging and merging, alerting, a health endpoint and shutting processes down cleanly

Syllabus

The path as it stands today: 9 modules totalling 11.4 hours. The certificate from June 2025 states 8.2 hours and 142 scrims — the path was extended after completion, recognisable by the modules marked “updated”. In the “Agents” module the multi-part ReAct and function-calling series are each condensed into a single entry.

Topics