For people who want to ship the system, not read about it
Getting an answer is easy. Getting it to stay up is the job.
12 classes from first prompt to deployed agent. Start tonight with 230 hours of recordings, or join the next live session on 25 Sep.
$250/mo recordings · $500/mo with live classes. Cancel any time. A class you finish stays yours.
Secure what you ship
$400 suggested
Own your deployment stack
$750 suggested
Add on anytime — keeps your current path. One-time purchase, classes yours to keep.
tms path build-ai-systems --stats
- classes
- 12
- recorded
- 230.5 h
- tools
- 42
A laptop, an API key, and some code.
This path assumes you can write code. The first class takes you from there. Not sure where you stand? Take the diagnostic — 2 minutes, no code required.
The bill nobody models first
What will this thing cost you to run?
The demo is free and the bill is not. Set your traffic and your token mix, put today's prices in the rate fields, and see where a rented GPU stops being the expensive option. Your inputs stay here: the bench runs in this browser, and the bill it reaches is saved in it.
What are you building? (sets sensible defaults)
A · Pay-per-token API
B · Serverless GPU (RunPod / Modal / Replicate)
C · Dedicated GPU (Lambda / RunPod reserved / your box)
Every rate is editable — pick a model or type your own numbers. Everything on the right recomputes from whatever you type. Tokens per second is real throughput, not the benchmark number. A month is 30.44 days.
See live benchmark data for every major model →Cheapest option at 100 requests a day
—
A · API
—B · Serverless
—C · Dedicated
—At your traffic and token mix, the dedicated GPU is the cheapest option. The API overtakes serverless at about 100 requests a day, and the dedicated GPU beats both at about 450 requests a day.
A serverless GPU charges only while it runs, so it wins at low traffic. A dedicated GPU charges all month, so it wins at high traffic. The API is the simplest but the most expensive per token.
Halve the prompt and the bill becomes — a month, — off, without changing vendor, model, or anything a user can see. That lever is the one nobody reaches for first, and pulling it is what most of this path is about.
Control AI Spending
Segment text into tokens and cost it. Run an open model on your own machine. The class that makes the number above yours instead of somebody else's.
Context Engineering
Design a recursive summarizer for documents several times the size of the context window — 329 exercises aimed squarely at the input-token half of that bill.
Production Agent Engineering
Configure an open model endpoint behind a provider interface, so switching between hosted and self-hosted is a config change rather than a rewrite.
Keep It Running
Produce an itemized bill for what your system costs, and work out what one user costs you to serve — including the user who costs several times the median.
The climb
What you walk out able to do
Every line below is something you can go and do afterwards, and the rung it sits on is how far into the work it is. Say how many hours you have and watch where they reach — or pick the classes yourself and see what they come to.
Spend it. Each class costs what it really takes.
- follow a model quickstart
- locate a tool in an open tool hub
- run a program someone else wrote and read its error
- execute a tool call round trip by hand
- locate a claim in its primary source
- locate the devices a passage uses on you
- run an open model locally
- configure an assistants memory and outside connections
- configure an open model endpoint behind a provider interface
- configure inference hyperparameters
- elicit output from a model
- express a data shape as a schema
- operate a model as a first pass editor
- produce a handoff that survives the author leaving
- produce a list of where untrusted input enters a system
- produce a moderation score for an output
12 more on this rung
- produce a refusal boundary that fires on cases you did not list
- produce a reusable system prompt
- produce a rubric a model can apply
- produce a stopping condition an agent can check on itself
- produce a usable interface for an agent
- produce an account of the gap between ideals and operating values
- produce an inventory of what you pay for and what it holds
- produce an itemised bill for what your running system costs
- recover a working state from your own version history
- segment text into tokens and cost it
- transform a corpus into an embedded index
- transform a transcript into a structured record
- characterise a models failure modes
- characterise a plans weaknesses with adversarial review
- characterise an applications injection surface
- characterise how a style works on a reader
- characterise how an AI assisted attack unfolded
- characterise what a tool costs you beyond its price
- characterise what one user costs you to serve
- characterise where a teams work stalls
- characterise who gains and who absorbs the cost of a deployment
12 more on this rung
- classify each recurring AI cost by keep downgrade or replace
- classify harmful outputs against a values statement
- classify models by fit for a workload
- classify where AI can fill a shape and where it must not choose one
- classify which framework applies where you are in the lifecycle
- constrain model output to a schema
- generalise a reasoning prompt pattern
- generalise a specification an agent can build from
- generalise a tool interface for model use
- generalise several sources into a position you can defend
- parameterise a prompt template
- parameterise an agent persona
- falsify a claim that a text was written by a person
- falsify a models alignment guardrails
- falsify a prompt with a benchmark
- falsify an applications defences against a named attack class
- hold several audience models on one surface
- justify a safety case for a system that acts without you
- justify a vector index for a latency and corpus budget
- justify an embedding model and its dimensionality
- measure a models value alignment against a stated standard
10 more on this rung
- measure whether a fine tune changed behaviour
- measure whether a user pays more than they cost
- verify a model output you thought was random
- verify a program does what you claimed with a test
- verify a rubric against independent graders
- verify a spending limit declines the charge you did not intend
- verify an agents actions with a critic
- verify model written code against its specification
- verify synthetic data preserves the property you need
- verify you would know your system broke before a user tells you
- construct a multi agent conversation with turn taking
- construct a predictive account of a models behaviour
- construct a retrieval system
- construct an agent that uses tools
- design a chunking strategy against measured retrieval
- design a packaged assistant over your documents
- design a recursive summariser for oversized documents
- design a review gate that catches what the doer cannot see
- design a system you can afford to keep running
2 more on this rung
- design an evaluation regime that decides model changes
- design an interaction protocol for responsible use
- arbitrate between the measure a system optimises and the goal
- arbitrate which internal feature drives a models behaviour
- govern a swarm by information hierarchy
- re architect an agent as a stateless reducer
- reconcile a failing retrieval by reranking and rewriting
- reconcile a working set with the context window it must fit
- reconcile agents from different toolchains into one run
- reconcile an agents run with a tool that failed
- reconcile what you built with what you can maintain alone
9 more on this rung
- select among agent architectures
- select among operating boundaries for inputs a system was not built for
- synthesise a long work that stays consistent across many generations
- synthesise a memory hierarchy for an agent
- synthesise a metalanguage for a problem domain
- synthesise a retrieval over relationships not just similarity
- synthesise a self extending agent behind a review gate
- synthesise an accountability regime for an unattended agent
- synthesise an agent that carries notes across its own runs
- originate an operating model for managed agent teams
- originate oversight that holds when the system outpaces the reviewer
3 of these 12 classes fit in 37 hours, and they open 44 of the 99 things above, as far up as write the playbook.
That opens the recordings, the exercises and the written curriculum on this path, and the standups every week. $500 a month adds a seat in every class on it that runs this month.
What a point is. Ten hours of your time — the recordings you watch, the exercises you work through, and the session itself. Move either slider, or pick the classes yourself and see what the hours come to.
Already on the calendar
The next 7 sessions on this path
The live plan is a seat in every one of them. Turn up with the traceback you are actually stuck on, ask about it out loud, and take the recording home afterward — it lands in the same login as the 230.5 hours that are already there.
tms schedule --path build-ai-systems
- 25Sep AI Alignment Friday · 1 exercise · $300 on its own next up
- 9Oct RAG & Memory Friday · $400 on its own
- 17Oct Control AI Spending Saturday · 3.1 h recorded · 12 exercises
- 23Oct Context Engineering Friday · 35.7 h recorded · 329 exercises · $400 on its own
- 26Oct Intro to Agents Monday · 56.7 h recorded · 32 exercises · $350 on its own
- 9Nov Agentic SDLC Monday · 26.5 h recorded · 46 exercises · $350 on its own
- 14Nov Using Large Language Models Saturday · 10.7 h recorded · 55 exercises · $60 on its own
Two ways in
Watch it all, or be in the room
Same curriculum either way. The difference is whether you are asking your questions out loud on the day, starting with AI Alignment on 25 September.
$250a month
Watch it all
- 230.5 hours of recorded sessions across 12 classes, yours immediately
- Exercises and the written curriculum for every class
- The tools built for these classes
- Pause, rewind, and run the exercise with the class on the other screen
$500a month
Be in the room
- Everything on the left, plus a seat in every upcoming class on this path
- Bring your own traceback and ask about it out loud — 7 sessions are already on the calendar
- Every session is recorded as it runs, so the hour you miss arrives days later
- The archive keeps growing while you are in it
Both renew every month and you cancel either one yourself, any time. Classes can also be taken one at a time, at their own prices — this path runs from $60 to $400 a class.
The sequence
12 classes, from prompt to deployed system
This order is curated, not alphabetical and not chronological — each class stands on the one above it. The bar under each class fills up as you build on what came before: purple is what you walked in with, cyan is what that class hands you.
-
01
Using Large Language Models
Class 1 of 12 -
02
Control AI Spending
Class 2 of 12 -
03
AI Alignment
Class 3 of 12 -
04
Context Engineering
Class 4 of 12 -
05
Claude Model Context Protocol
Class 5 of 12 -
06
Intro to Agents
Class 6 of 12 -
07
Prompt Engineering
Class 7 of 12 -
08
RAG & Memory
Class 8 of 12 -
09
Advanced Retrieval Augmented Generation
Class 9 of 12 -
10
Production Agent Engineering
Class 10 of 12 -
11
Agentic SDLC
Class 11 of 12 -
12
Agentic AI Security: Securing What You Build
Class 12 of 12
Straight answers
What is behind the login
230.5 hours of this path, recorded
Prompt Engineering is 63.5 hours; Intro to Agents is 56.7 hours; Context Engineering is 35.7 hours. 230.5 hours across 12 classes.
Pause it, rewind the part where the trace does not match the code, and run the exercise with the class still on the other screen. Start any week and the sequence picks you up.
And the room they were recorded in
A recording answers the question the teacher expected. The room answers the one you brought about your own repo. The next one is AI Alignment on Friday 25 September — the live plan is a seat in it.
Every session is recorded as it runs, so the hour you could not make arrives in the same login a few days later, and the archive you joined keeps getting longer while you are in it.
Both plans, and what carries on
The recordings, the exercises, the tools and the written curriculum come with either one; the dearer one adds the live room.
Watch it all $250 Be in the room $500The tools that come with the classes
You open these next to your own work — your prompt, your repo, your bill — with the recording paused on the other screen.
- Using Large Language Models 4 guided workbenchs, 3 aids.
- Control AI Spending 1 guided workbench.
- AI Alignment 1 guided workbench.
- Context Engineering 4 guided workbenchs, 3 aids. Nine stations that walk you from a bloated prompt to a context budget you can defend, on your own material.
- Claude Model Context Protocol 1 guided workbench.
- Intro to Agents 2 guided workbenchs, 1 aid.
- Prompt Engineering 3 guided workbenchs, 1 aid.
- Production Agent Engineering 1 guided workbench, 1 aid.
- Agentic SDLC 3 guided workbenchs, 1 companion, 1 aid, 9 walkthrough decks. A companion for running the loop, plus decks on agent memory, the complexity ladder, context compression, git under agents, and shipping.
- Agentic AI Security: Securing What You Build 1 walkthrough deck, 1 reference. A frameworks reference and a map of where untrusted input gets into the thing you built.
What to expect
What this path asks of you
- You will write code and live in a terminal. Not "a bit of Python eventually" — from Intro to Agents onward you are running processes, reading stack traces and configuring endpoints yourself. If that sentence is the appeal, you are on the right page.
- You end up owning the running system, including the bill, the outage and the injection surface. That is the point: nobody can take it away from you afterward.
- 230.5 hours across 12 classes. This path moves fast and builds in sequence.
- Curated, not a buffet. Both plans are 12 classes deep and one path wide, in a deliberate order, so each one stands on the last. Want to browse first? See all learning paths.
And if you want AI doing your work without building the plumbing yourself, two other paths cover the same ground with no terminal in them:
One sequence, worked out already
Press play tonight, be in the room on 25 September
230.5 hours of recorded sessions open the moment you join, in the order the school teaches them, and every class above tells you what you will be able to do — and what you hand over to prove it — before you spend an hour on it. That next session is AI Alignment, and the live plan is a seat in it.
The standups run every week on both plans. Both renew every month and you cancel yourself, any time.