Make an AI assistant ready for real customers
You start with an assistant that answers well on a normal day and breaks on a bad one: too slow, too expensive, failing when the model provider does, open to attacks, impossible to watch. You write the parts that fix that, and a scoreboard shows each fix turning a red line green. It runs on your own computer, with Claude Code or your own Claude or OpenAI key.
- Guided pages for all ten modules, read online
- The project kit to download: plain Python, runs on your computer
- A worked example to compare your results with
- The track's paid lessons
Python 3.10 or later, and Claude Code, a Claude key or an OpenAI key.
Start with setupPurchases are available in the US for now.
What you end up with
- A model chosen by a test you wrote, including a test for a fact hidden in the middle of a very long document.
- Saved answers that are safe (never one customer's order shown to another), a trimmer and a sorter that send easy questions to a cheaper model.
- Search that finds better pages for vague questions, with scores that prove the gain is real and not luck.
- A multi-step job that survives a crash and never does a refund twice.
- Retries, a rate limit, a switch that stops calling a failing service, and a backup model, tested against failures injected on purpose.
- A trace you can follow for any request, speed and cost numbers, a canary check for new versions, and a drift alert.
- Defences against tricks in messages, in documents and in replies, and a rehearsed bad day with a postmortem and a launch checklist.
What you need
- Python 3.10 or later, and a terminal. You will write Python: functions of 10 to 40 lines, each described in plain words.
- The Beginner project, or comparable experience building an assistant that answers from documents. The assistant you start from is supplied.
- One model, your choice: Claude Code (signed in, no key to enter), your own Claude API key, or your own OpenAI API key. A second, cheaper model is optional.
- About 30 to 40 hours, over as many sittings as you like. Your progress is saved.
- The checks on your code are free and instant. The live checks cost roughly $1 to $3 in model calls at API prices; with Claude Code they count against your plan instead. The kit shows an estimate and asks before it spends anything.
How a module works
- Module 0 shows a scoreboard, red, from a practice day replayed on a simulated model. Each line is one real-world worry, with a bar to clear.
- Each other module tells you which file in
svc/to open. The file explains each function you must write. - You write it, then run the module. It first runs checks on your code. They use a fake model and a clock that skips waiting, so a retry that would take a minute is tested in milliseconds, and each check says in plain words what it expected and what it got.
- When every check passes, the module tries your code against your real model, and counts the module as done when the numbers clear the bar. Run module 0 again at any time to see your scoreboard move.
The ten modules
Each module is one guided page and one run of the project. The pages are part of the track: they open once you own it.
- Module 0 See what breaks: run the starting assistant through a bad day
- Module 1 Choose the right model by testing models on your own questions
- Module 2 Make it cheaper and faster by reusing answers, sending less text and using a cheaper model for easy questions
- Module 3 Decide whether to retrain the model or fix the problem another way, and prepare clean training data
- Module 4 Make search find better pages, and prove it is really better and not luck
- Module 5 Make a multi-step job finish even if the program crashes halfway, without doing a step twice
- Module 6 Keep working when the model fails: wait and try again, stop hammering a failing service, and switch to a backup
- Module 7 See what is happening in production: follow one request step by step, track speed and cost, and test a new version on a few customers first
- Module 8 Stop attacks: block tricks in customer messages and documents, and stop private data leaving through links
- Module 9 Rehearse a very bad day, then write what went wrong and a checklist before launch
Also in the track
Setup, troubleshooting, a worked example to compare your results with, and a guide to running a model on your own computer with Ollama.