Writeups
How the AI systems on this site actually work, and the reasoning behind them.
A Self-Hosted Chatbot That Won't Fake a Tool Result
A chatbot on my old gaming laptop, eleven tools, and one rule: if it says a tool told it something, a tool had to have actually told it that.
I Gave a Fruit Fly's Real Brain a Tinder Account
783 real neurons and 8,454 real synapses from a published Drosophila connectome, wired up to judge dating and food photos. No invented circuitry, just a real brain given a new job.
An MCP Server for ATS Resume Scoring, With No UI
One tool, analyze_resume, meant to be called by an agent, not a person. The deterministic/LLM scoring split, and how it stays prompt-injection safe.
Someone asked Jinnah how big his pp was. Here is what a month of real questions taught me.
805 real conversations, 16 countries, and one bug I only found because I sat down and actually read the logs.
Driving a Unity Humanoid Avatar with MediaPipe, No Depth Camera Required
Retargeting MediaPipe’s 3D pose landmarks onto a Humanoid Animator with Quaternion.FromToRotation, running entirely on-device through Unity Sentis. The approach, the calibration trap, and setup steps.
The Numbers in a Private Equity Deal Get Tested Systematically. The Reasoning Never Has.
Private equity has spent decades scrutinizing the model and waving the argument underneath it through. Large language models are the first thing capable of testing that reasoning with the same rigor.
I built an AI that answers as Muhammad Ali Jinnah, using only his own words.
Most of the work went into making sure it can't lie about him.