A book in progress
AI Systems Engineering
From your first API call to making autonomous agents. The work between a demo that runs once and a system you'd let a customer touch.
I. The thesis
The demo was never the hard part
Wiring up an LLM and getting an impressive answer back takes an afternoon. Getting that same thing to behave when real users hit it, when the input is messy, when it has to retrieve the right document, when it can't quietly make something up, is where most projects stall. MIT's 2025 study of enterprise GenAI put numbers on it: the large majority of pilots never produced a measurable result.
The gap isn't the model. It's everything between a working demo and something you'd let a customer touch.
This book walks that gap in order. It starts at a first API call and builds up: prompting that's actually reliable, retrieval that fetches the right thing, evaluation so you know when a change made things worse, and agents that can act without going off the rails. Every chapter is something you build and run, not a diagram you nod at.
It names specific tools and is honest about their trade-offs, including where they fall short. The aim is that an engineer who finishes it can take an AI feature from idea to production and defend the decisions they made along the way.
II. Who it's for
Two readers in particular
Engineers new to AI
Backend and full-stack developers who know how to ship software but haven't built with LLMs yet, and want a straight path in without wading through hype or maths they don't need.
Builders stuck at the demo
People who can already call an API and get a result, but are struggling to make it reliable, testable, and safe enough to put in front of real users. The book is mostly about that second mile.
III. What it won't be
A short list of what this book avoids
- A from-scratch tour of transformer internals. Plenty of books do that already, and you don't need it to build well.
- Framework tourism. The book commits to a small set of tools rather than chasing the library of the week.
- Toy demos that only work in a notebook and quietly fall apart the moment a real input arrives.
- Predictions about AGI, or claims that agents are about to replace the people reading this.
- Vendor-neutral hedging. Where one approach is better for a job, the book says so and explains why.
IV. Author
Why me
I'm Tamas Piros. Google Developer Expert in Web Technologies. 25+ years building software, and a long stretch teaching engineers, through training I've run and conference workshops, with tens of thousands of developers along the way.
I build with AI daily, across LLMs, vector search, agents, MCP servers, evaluation harnesses, and on-device inference with Gemma. The chapters come from work I've actually shipped and debugged, not from reading other people's posts.
This is the book I keep wishing existed when an engineer asks me where to start and how to get past the demo. So I'm writing it.
V. Keep me posted
Email me and I'll let you know when the first chapters are ready, and when the book launches.
I'll also send the occasional short note on where the book's going. No frequent newsletter, no promotion for anything unrelated.