NestedQuotes Logo
Profile photo of Jarrod Connolly

Jarrod Connolly

Writing at NestedQuotes

From Node to Narnia, from pixels to Python, it was quotes all the way down.

Follow
Recent Articles

suBPEriod - Byte Pair Encoding at speed

September 21, 2026

suBPEriod - Byte Pair Encoding at speed

C++   Performance   Tokenizers  

suBPEriod is a C++ Byte Pair Encoding encoder I wrote so I could work on speed without giving up accuracy. Unit tests require the same token ids as tiktoken, including Unicode, and a fuzzer checks that odd inputs do not hang or crash. I kept a change only when a second run and the profiles agreed. On a 100 megabyte English sample it used about half of tiktoken's time, and Chinese stayed near parity.

Tutorial Videos - Automated

August 21, 2026

Tutorial Videos - Automated

PAC Hub  

Product tutorial videos go stale the moment the UI changes. This article shows how I generate PAC Hub training videos with Playwright and Grok TTS, and why keeping them current costs pennies. The pipeline records the real product, speaks a script, and can be run again when the interface moves. A script is the source of truth, so I can regenerate a video instead of reshooting it.

Why I Am Building PAC Hub

July 1, 2026

Why I Am Building PAC Hub

PAC Hub  

I chair a Parent Advisory Council at a large elementary school, and the tools around that work add a lot of extra weight. I am building PAC Hub so parent volunteers can run a school year from one place. The aim is less toil, a cleaner handoff, and a home the next chair can inherit.

Chunking Technique Research

June 9, 2026

Chunking Technique Research

AI   Python  

After building RAG systems for technical books, I kept asking which chunking strategy actually retrieves better. I built a research platform to answer that with real vector search, LLM-as-judge ground truth, and metrics such as NDCG, MAP, and MRR. I compare fixed-length and overlap splitters with heading-aware and linguistically scored approaches. Some of this is still a heuristic, not a settled result.

Consuming Textbooks

May 28, 2026

Consuming Textbooks

AI   Learning  

I never finished technical textbooks by reading them cover to cover. This article describes how I use an LLM as a chapter-by-chapter study partner, with room for side quests and real code along the way. The point is not to hand off the understanding. It is to keep the doing in the loop so the material sticks.