Decide,Don'tGenerate
Jev, System One Models, and the Decision Layer of Agentic AI
Most of what an AI agent does isn't writing. It's deciding.
← Drag to turn the book →
Read it the way you like to read.
The Amazon editions are on their way. Leave your email and we'll write once, the day they're live.
Is this alert real? Which team gets this ticket? Is this action safe? Those are decisions, not writing.
Most agents give every one of those small decisions to a large language model, then search its paragraph for the answer. It works, but it is slow and expensive, and the model sounds just as sure when it is wrong.
This book shows you a better way to build the decision layer: calibrated probabilities that mean what they say, thresholds set by what each mistake costs, and a clear rule for when to give a case to a person. Every step is built with runnable code, around one realistic (synthetic) security team at a company called Kestrel Logistics.
- 01Check whether a model's probabilities mean what they say, and fix them when they don't.
- 02Turn a probability into act, review or escalate, with thresholds set by real costs and real capacity.
- 03Compare six ways to make the same decision, from hand-written rules to LLMs to Jev.
- 04Build a hybrid agent in which Jev decides and the LLM reads and writes.
- 05Build your own small System One model, with typed heads and calibration built in.
Turn the real pages.
Real pages from the book. Tap any page to read it.
Let the costs draw the lines.
Write down what each action costs and the thresholds draw themselves. Move the sliders to see Kestrel's lines move.
Analysts miss about one threat in twenty, as in the book. Kestrel's numbers are the author's estimates, not industry figures.
Expected cost of each action as P(real threat) rises, both axes on a log scale. The crossings are the thresholds. Synthetic, as in the book.
Seven parts, from probability to production.
Three ways through the book.
You don't have to read it front to back. Pick the route that starts where you are.
Read it in order.
Part I builds the foundation everything else rests on. Parts II and III explain networks, embeddings, attention, LLMs and agents without assuming you've met them before.
Part I, then jump to Part IV.
Don't skip Chapter 4 on calibration or Chapter 5 on costs. The rest of the book leans on those two chapters more than on anything else.
The questions to ask.
Part I, then Chapters 15 to 21, then 25 and 29. Skip the code and still ask the right questions: what does a mistake cost, where's the line, who reads the flags?
Sridhar Mukkandi
An applied AI engineer who builds AI agents and retrieval systems, leads and trains a team of 15 engineers in generative AI, and has shipped AI systems for government and enterprise clients. He wrote this book for engineers who want AI systems they can trust, and founded Tarkika to keep writing them. More than 280,000 people follow him on Instagram and YouTube.
A book you can run.
Before you buy.
Do I need an API key or a paid account to run the code?
No. The book's toolkit, jevkit, includes a free mock that answers the official SDK's calls without a network connection or a key. If you have a real key, one setting switches the labs to the live service.
How much maths and code do I need?
A little arithmetic and a little Python. Part I starts from the beginning: what learning means and what a probability is. If you already build with LLMs, you'll move quickly through Parts II and III.
Is this book about one product?
No. Jev is the running example, but the idea behind it is worth understanding whatever happens to any one product: some decisions belong in a separate, fast, measurable layer of an AI system. It's an independent guide, not affiliated with TypeSafe AI, and every Jev number in it is marked synthetic.
Can I read it on Kindle in India?
Yes. The Kindle edition is available on Amazon.in and works on any phone, tablet or Kindle through the free Kindle app.
How fast is delivery in India, and can I pay on delivery?
Paperbacks ordered directly ship within 48 hours. Most cities receive them in 2 to 6 working days. There is nothing to pay online: you pay the courier in cash or with any UPI app when the book arrives. Delivery is free, and you get updates on WhatsApp. A copy damaged in transit is replaced free.
Do you offer bulk copies for colleges and teams?
Yes, with discounts for classes and teams, and an optional workshop built on the book. Write to hello@tarkika.com with the number of copies and the city.