AI
Engineering
The real stack behind production AI products: prompting, retrieval, agents, tool use, structured outputs, and evaluation, applied to a real shipped product, not a toy chatbot.
9 modules, in order.
Orientation first, then four stages that each build on the last, ending in a capstone graded against a rubric, not vibes.
Orientation
What this field actually is, what people in it actually build, where you might fit, and exactly what the next 8 modules will (and won’t) teach you.
Prompting & Structured Outputs
A language model produces text. Your program needs data. This module closes that gap reliably, not with a nicer-sounding prompt, but with validation and repair logic that doesn’t trust the model to behave.
Context Windows & Token Economics
Every prompt has a budget, measured in tokens, not characters or “how much text feels reasonable.” The actual arithmetic for what fits, what it costs, and what to do when your content doesn’t fit.
Retrieval-Augmented Generation
A model only knows what’s in its prompt. Retrieval decides, out of everything you could put there, which pieces actually get in, by searching, not by hoping the right chunk made the cut.
Tool Use & Function Calling
A model can only produce text. Everything it does beyond that, looking something up, sending an email, charging a card, happens because a tool call it generated got executed by your code. That execution step is where correctness and safety actually live.
Agent Architectures & Control Flow
An agent is what happens when a model chains steps together, reasoning about each result before deciding the next action, in a loop it controls. That loop needs boundaries, or it doesn’t stop on its own.
Reliability for LLM Systems
Model APIs fail: rate limits, timeouts, transient server errors. Some failures are worth retrying, some never will succeed no matter how many times you try, and treating them the same wastes time, money, and your users’ patience.
Evaluation, Testing & Observability
“I tried a few prompts and it looked good” isn’t evaluation, it’s a vibe check on a handful of lucky examples. A systematic way to measure whether a system is actually working, and whether a change made it better or worse.
Capstone: A Production AI Agent
A client brief: a support bot that must answer from a knowledge base, act on tools safely, run inside a bounded loop, and survive a flaky model API, graded against a rubric, not vibes.
- Answers from a knowledge base
- Acts on tools safely
- Runs in a bounded loop
- Survives a flaky model API
- Graded against a rubric
Final module: 5 AI agents you can sell
The organized code for each agent, the best model for each, who to sell it to and where, the cold emails to send, how to run clients and work out your revenue, and a 30-day plan to your first paying client.
- Support Agent
- WhatsApp Lead Agent
- Inbox Manager
- Booking Agent
- Document Extractor
Finish the course with 5 AI agents you can start selling.
The final module hands you their organized code, plus who to sell each one to and how.
- Organized Python code for all 5 agents, with tests
- The best model for each agent, with verified prices
- Niches and cities to sell each one to
- Cold email templates and the rules to follow
- A client guide and tracker, a revenue guide and calculator, and a 30-day sell plan
Support Agent
Answers customer questions from the client's own knowledge base, or hands off.
Model, niches, cities and emails unlock inside the courseWhatsApp Lead Agent
Replies to new enquiries in seconds, qualifies them, and passes hot leads to a person.
Model, niches, cities and emails unlock inside the courseInbox Manager
Triages email and prepares reply drafts for a human to review. It never sends.
Model, niches, cities and emails unlock inside the courseBooking Agent
A bounded tool loop that finds a slot and books it once.
Model, niches, cities and emails unlock inside the courseDocument Extractor
PDF in, validated JSON out, flagged for review when the numbers do not add up.
Model, niches, cities and emails unlock inside the course
Starting prices and niche lists inside the course are suggestions to test and research, not promises of income.
5 sections, 16 real pages.
The course is an app, not a video playlist. Every page has one job, from where you left off to proof you can show.
Learn
5 pagesCourse structure, reference material, and where you left off
Practice
2 pagesExtra reps, quizzes, and skill tracking
Build
1 pageThe real projects and labs you’ll actually ship
Evidence
3 pagesPortfolio, career path, and proof you can show
Final module
5 pagesFive sellable agents, their code, and how to sell them
Skills you ship, not slides you watch.
Six things you can do by the end, each one practised on a real product rather than a toy example.
The gap between a demo and a product.
- 01Retrieval can ground an LLM on information outside its training data.
- 02Tool use lets an AI system interact with external systems, not just generate text.
- 03Evaluation is necessary because a response that sounds right can still be wrong.
- 04Prompting is only one part of a production AI system.
- 05Cost and latency become real engineering constraints once an AI application scales.
A shipped AI product with evals you can point to in an interview.
Whether you’re moving into AI engineering freelance or hired.