Blog

Folders make you decide where something belongs before you know where you'll need it. I made that bet in a side project and it broke the whole point of it.

I found a tiny age penalty in my ranking code: 5% a day. After six months it scored a perfect match at 0.0001. The maths, and why my tests never noticed.

LLMs don't remember anything. "Memory" is text an app quietly pastes into your prompt. How that works, why it stays stuck in one app, and what I measured.

Events are billed per event, per consumer, and per kilobyte. I did the math at three volumes so your AWS invoice doesn't have to surprise you.

Split a monolith and every multi-step operation becomes a distributed transaction. Here's why I'd hand that job to an orchestrator.

A model never sees your text. It sees a list of integers called tokens, and that list is what you pay for and what your context limit counts.

Semantic search matches meaning, not keywords. The embedding model was the easy bit. Picking a vector size and an index my database could serve was the job.

Usually not. Most LLM features are cheaper, faster, and easier to trust as a fixed workflow. Here's when an agent earns its keep.

MCP is a standard plug for AI tools. Write a tool once and any app that speaks the protocol can use it. What it fixes, and what it costs you.

An LLM just predicts the next token. Everything useful is the scaffolding around it: context, tools, and memory. My map, via LangChain and LangGraph.

A JSON parser is two small programs: a tokenizer and a recursive descent parser. I built both in TypeScript on Deno to see how they tick.

A spell checker is edit distance plus a dictionary. Here's how Levenshtein distance works, with code and a visualizer that shows the matrix filling in.

JavaScript registers every declaration before it runs any code. That one fact explains undefined from var, callable functions, and the temporal dead zone.

Every bit of JavaScript runs in an execution context built in two passes: memory first, then code. Trace that by hand and the weird stuff stops being weird.

Big O is just how much more work your code does when the input grows. Four common shapes, JavaScript examples, and what they cost at real sizes.