B-Trees or LSM-Trees: Your Database Already Chose
Every database picks a side in the same trade — quick to write, or quick to read. Almost every performance surprise you have had follows from which side it picked.
derpx06Notes on systems, models & learningWriting
11 pieces on the systems I build, the models I break and the way I learn. Newest first is rarely the best order — start from a tag.
2026
Mobile Apps Are Distributed Systems With Terrible Networks
A phone is a copy of your data that vanishes without warning, holds changes for hours, and comes back disagreeing with everyone else. Most mobile bugs are really that problem, unlabelled.
What I Got Wrong About Document Modelling
Picking MongoDB takes five minutes. Deciding what one record should contain takes months to get wrong and is painful to undo.
Your Retry Made It Worse
Trying again after a failure is the most obvious safety net there is, and the easiest one to build backwards. It hinges on one question: did the first attempt actually fail?
MCP Is a Protocol, Not a Plugin System
The Model Context Protocol looks like a way to hand an AI some plugins. Treat it that way and you get tools that shine in a demo and go round in circles in real use.
Local-First AI Is a Systems Problem, Not a Model Problem
Getting an AI running on your own laptop takes four minutes. Everything that makes it actually usable afterwards is queueing, memory and patience — plain old engineering.
Evals Are the Only Moat
Anyone can copy your prompts in a weekend. Nobody can copy the thing that tells you, precisely, whether your last change made the product better.
Your RAG System Doesn't Have a Retrieval Problem
When an AI answers badly from your own documents, everyone blames the search. In every system I have debugged, the search was fine. The documents were the problem.
I Fine-Tuned a 0.5B Model to Write Shell Commands
A tiny AI running on my laptop beats a giant one at a single job. That sounds like a story about size. It is really a story about picking a small enough job.
Inside Claude's J-Space: Anthropic's Global Workspace Research
Anthropic found a small, editable workspace inside Claude that carries silent intermediate thoughts. Here is what the J-lens reveals, why it matters, and what it does not prove.
KV Cache from First Principles
A clear first-principles guide to KV cache: why it exists, how it works, and what tradeoffs matter in real production systems.
Start a course
Finetuning an LLM
Start lesson 1The monthly letter
One email a month
What I read, built and got wrong.