Prompting is a tutorial. Composition is the trade. Learn to assemble context, retrieval, agents and evaluation into one production system — and practice every lesson inside a real, live one: Alice.
Every production AI system — including the one this course runs on — is a composition. Data boots up, gets retrieved, an agent loops over it with tools, state carries the memory, and evaluation catches what it misses. This is the machine you will learn to build.
Every module ends with your hands on the machine. Four of them have live labs on this page right now — the same labs, bigger, wait inside the course.
Tokens, context windows, and why the machine sounds certain when it's guessing. The one mental model that explains everything AI does — why it forgets, why it invents, why order matters. Start here, and nothing downstream ever looks like magic again.
Open the Context Lab →Before an agent thinks, it loads. Business facts, prices, rules, history, guardrails — stacked in the right order, every single time. Get the boot wrong and everything downstream is fiction.
Open the Boot Lab →The machine forgets everything the moment a chat ends — unless you build it a memory. Working memory, long-term memory, and state: how a system holds a customer's name, last Tuesday's quote, and where a job stands — across days, not seconds.
Not "upload a PDF and pray." Pre-retrieval, retrieval, post-retrieval — filtering, scoring, re-ranking — so your agent answers from the right page of your business, not a plausible one.
Tools, tool tracking, state. The loop that turns a chat into a worker: answer, quote, book, follow up — in the right order, with a human safety net woven in.
Open the Playground →Chat is only one sense. Give your system ears and eyes: phone calls it can hold, photos it can read, documents it can understand — the same senses running live today in Alice's phone receptionist and photo estimator.
Train the system on your trade, your prices, your voice — no code. The owner-level version of fine-tuning: examples, corrections, and a knowledge bank that compounds.
Production systems are graded, not trusted. Score transcripts, hunt the planted misses — invented prices, dropped bookings, broken guardrails — and fix them before customers find them.
Play Miss Hunt →Power without a fence is a liability. What the machine must never say, never store, never send — and when local beats cloud. The module that keeps your system, your customers, and your good name safe.
No certificate for watching. You graduate when one production system — yours — is live, taking real requests from real customers, with you at the console. The capstone ships two things: the real system, and 🌐 your own developer portfolio site — the page that presents it to employers. Pass, and the 🏆 certificate of completion is awarded.
Nine advanced modules, opening to Composer graduates first. This is the second tier: where builders turn a working system into an unfair advantage — and an offer.
Cohort 2 contains nine advanced modules.Hover or tap a module — what you'll learn is inside.
AI that books, charges, and messages in the real world, on a leash you design.
What every call costs, and when the small cheap model beats the giant.
A full module of broken systems. Diagnose the failure before you're allowed to fix it.
Your prices, reviews, and job history become the training set.
Package what you built: price it, demo it, guarantee it.
Instructions an agent loads only when the job needs them. A lean core in its head, a deep bench on the shelf — and no context wasted carrying all of it.
One standard plug between your agent and the world. Build the connection once, and any tool, database, or prompt snaps into it.
Many agents, one conductor. You score who does what, in which order — then the system plays it back the same way every time.
Agents that talk to other agents across systems — delegating whole jobs machine-to-machine and getting finished work back.
Cohort 2 enrollment opens to the waitlist after Cohort 1 ships.
Four working slices of the course, live on this page. Fill the window. Toggle the context. Assemble the pipeline. Hunt the misses. This is how every module in THE COMPOSER feels.
Everything the machine "knows" during a conversation lives in one window, measured in tokens — roughly one word each. This demo window holds 120. When it overflows, the oldest words fall out the back and stop existing. That's not the machine ignoring you; that's why it "forgot" your name. Fill it and watch.
One customer question. One agent. Five context layers you control. Flip layers off and watch the exact same agent get dumber — or dangerous. This is why the boot sequence is Module 1: everything an agent says is downstream of what it loaded.
Click blocks to assemble your agent's pipeline, in order. Then run a live request through it. Order is not decoration — it's the difference between a booked job and a booked headache. Click a placed block's ✕ to remove it.
Three real-style transcripts. Each hides exactly one miss — an invented price, a dropped booking, a broken guardrail. Click the message you believe is the miss. Evaluators don't trust systems. They grade them.
The labs on this page run on the same ideas that power Alice AI Office Hub — the course feeds the system, and the system feeds your business.
No checkout yet, no countdown theater. Enrollment opens soon — the waitlist is the door, and founding pricing dies when the first cohort fills.