- Services
-
Case Studies
- Careers
-
Blog
-
When to Call the LLM?
- May 27, 2026
Refactoring the Book2Audio Parsers
- Apr 30, 2026
-
- Generate Password
You told the LLM to fix OCR artifacts. It also "fixed" words that weren't broken. "Judiciary" became "judicial." Em-dashes quietly became hyphens. A numbered list item lost its prefix without a trace. None of it was malicious — the LLM just can't help optimizing toward well-edited when you asked for OCR-cleaned. Those aren't the same thing. The solution: a guard layer that watches what the LLM changes and silently undoes anything it had no business touching. Sounds simple, right? Then the edge cases and failing unit tests really begin.
My AI pipeline was slower than it needed to be. When building a Book-to-Audio converter, the obvious culprit was the LLM cleaning every paragraph of OCR-mangled text. The fix seemed simple: add a gate that skips clean paragraphs entirely. It worked. But then the timing data revealed a second hidden slowdown that the optimization itself had exposed — one that was invisible until the gate made it visible.
What if you could write a program without actually writing it? Genetic Programming does exactly that — it evolves code the same way nature evolves species: randomly, ruthlessly, and surprisingly effectively. In this post I walk through how it works, build up a toy version from scratch, and then turn it loose on a hidden mathematical function to see if it can find the answer on its own. Spoiler: it can. But the more interesting question is how.
There's a mathematical proof that says no algorithm — no matter how clever, how sophisticated, or how well-designed — can outperform random guessing when averaged across all possible problems. Not A*, not neural networks, not humans in the loop. Every advantage on one class of problems is paid for, dollar for dollar, somewhere else. It sounds like it should be false. It isn't. And once you understand why, you'll never look at machine learning the same way again.
What does it actually take to plug a brand-new AI voice engine into an existing codebase without breaking everything else? This post pulls back the curtain on the code behind Book2Audio's Qwen3-TTS integration — from the two-method abstraction that makes any TTS engine swappable, to the GPU memory tricks that squeeze a 1.7B model onto a consumer laptop.
The unglamorous work of software is making things consistent — and this update to Book2Audio is exactly that. Two parsers that did the same job differently have been brought into alignment, sharing a single text-cleaning pipeline and a unified paragraph accumulation strategy. It's the kind of refactor that doesn't change what the tool does today, but makes possible something much more interesting tomorrow: an LLM-based cleaning step that can fix the OCR errors, broken page splits, and stray footnotes that rule-based cleaning can never quite catch.
Every learning algorithm is making a bet. It can't prove its predictions from the data alone — it's sneaking in assumptions, whether it admits them or not. Name those assumptions precisely enough, and something surprising emerges: there's no such thing as induction. It's deduction in disguise. This post unpacks what that means, why stronger assumptions lead to better generalization and more spectacular failures, and what it reveals about neural networks that most people never think to ask.
Book2Audio started as a PDF converter — but the best books often come as EPUBs. This post walks through how we added EPUB support, what it took to strip a RAG-focused parser down to something clean enough for audio, and the small details that turn out to matter when you're reading a book aloud: chapter titles that actually get spoken, footnote markers that don't interrupt the flow, and a debug mode that shows you exactly what the parser heard before you commit to a three-hour conversion run.
What if you could turn any PDF into an audiobook and fine-tune the narrator's voice with nothing but a plain English instruction? Alibaba just open-sourced Qwen3-TTS, and it's worth paying attention to — nine built-in speakers, natural language style control, and weights you can run on a laptop GPU. But does it actually sound good enough to listen to for hours? The answer might surprise you.
Machine learning textbooks have their own vocabulary. But behind the jargon lies a process that would be deeply familiar to Karl Popper: conjecture and refutation. This post is a short reference guide
