I have a problem that I suspect a lot of engineers share.
I buy technical books because I want to understand things properly. Distributed systems, event-driven architecture, domain-driven design, database internals, cloud infrastructure. My library keeps growing, and I genuinely enjoy building it.
But there’s a catch.
When I’m working on a problem, I rarely remember exactly where I read something. I might remember that an author explained idempotent consumers particularly well, or that a specific chapter finally made consensus algorithms click.
I know the information is somewhere in my library. Finding it is another story.
Opening PDFs, searching through chapters, jumping between books, and trying to remember which author explained something best is not how I want to spend my engineering time.
I wanted something better.
I wanted to open a terminal, ask a question in natural language, and have an AI assistant search my actual books, understand the relevant passages, and give me an answer with references to the original material.
So I designed and built a personal AI book assistant on AWS.
In this article, I’ll walk through three things:
The architecture: How I built the system on AWS, from document ingestion and indexing to AI-powered retrieval and answers.
What I learned building this: The challenges, costs, and engineering decisions worth knowing.
My final takeaways: What worked, what I’d do differently, and where I’d take the project next.
Let’s get into it.
1. The architecture: how I built the AI book assistant
Before getting into the implementation, let’s look at the architecture I ended up with.
I wanted two things from this system.
First, I needed a reliable way to turn my collection of technical books into a searchable knowledge base without manually processing documents every time I added something.
Second, I needed an intelligent interface that could understand my questions, decide when it needed additional information, retrieve the relevant passages, and generate answers grounded in my books.
These requirements led me to separate the solution into two independent workflows: ingestion and query.
The ingestion workflow runs whenever my library changes. The query workflow runs whenever I interact with the assistant.
They share the knowledge base, but otherwise have separate responsibilities.

