https://codex-crusader.github.io/writing/ Posts on local-first software, machine learning, statistics, security and UX research. en-in Sat, 03 Oct 2026 15:30:00 +0000 https://codex-crusader.github.io/writing/local-ai-desktop-companion-hyprland/https://codex-crusader.github.io/writing/local-ai-desktop-companion-hyprland/Sat, 03 Oct 2026 15:30:00 +0000Rice-sauce-hair is an Arch Linux and Hyprland desktop with a local AI companion on Ollama. How she turns vague words into actions, and why code, not the model, decides what is safe.Software engineeringMachine learning https://codex-crusader.github.io/writing/correlation-engine-noise-baseline/https://codex-crusader.github.io/writing/correlation-engine-noise-baseline/Tue, 29 Sep 2026 09:00:00 +0000The Correlation Engine runs about 4,875 hypothesis tests a day across news, markets and government data. How it separates real patterns from chance, and why it shows the noise result next to the real one.Data https://codex-crusader.github.io/writing/data-leakage-basketball-model/https://codex-crusader.github.io/writing/data-leakage-basketball-model/Tue, 29 Sep 2026 09:00:00 +0000An NCAA basketball predictor reached an AUC of 0.9666 because it trained on stats from the games it was predicting. How the leak happened, how I found it, and the checks that stop it now.Machine learningData https://codex-crusader.github.io/writing/alphazero-chess-engine-python/https://codex-crusader.github.io/writing/alphazero-chess-engine-python/Tue, 29 Sep 2026 09:00:00 +0000AZ-Lite is a small chess engine that learns only from self-play. How its tree search, policy and value network, and training loop fit together, and what it does not do yet.Machine learning https://codex-crusader.github.io/writing/campus-visitor-access-ux-study/https://codex-crusader.github.io/writing/campus-visitor-access-ux-study/Tue, 29 Sep 2026 09:00:00 +0000A UX case study of visitor entry at Vijaybhoomi University: interviews, card sorts, tree tests, a Figma prototype, eight usability tests, and a working app rebuilt from the results.Design https://codex-crusader.github.io/writing/local-first-software/https://codex-crusader.github.io/writing/local-first-software/Fri, 28 Aug 2026 09:00:00 +0000Most software you use today runs on someone else’s computer. Local-first software runs on yours. Here is what that changes, what it costs, and why nearly everything I build works this way.Software engineeringSecurity https://codex-crusader.github.io/writing/explainable-market-intelligence/https://codex-crusader.github.io/writing/explainable-market-intelligence/Fri, 28 Aug 2026 09:00:00 +0000The Pulse Engine reads prices and news and prints one number. The number was the easy part. Making it say how it got there took far longer, and that is the part that makes it worth anything.DataMachine learning https://codex-crusader.github.io/writing/transformer-from-scratch/https://codex-crusader.github.io/writing/transformer-from-scratch/Fri, 28 Aug 2026 09:00:00 +0000I built a transformer from nothing, starting at the character tokenizer and working up through BPE and RoPE. Not to produce anything useful, but to find out whether I understood the internals or only the diagrams.Machine learning https://codex-crusader.github.io/writing/neuroevolution/https://codex-crusader.github.io/writing/neuroevolution/Fri, 28 Aug 2026 09:00:00 +0000Almost every neural network is trained by backpropagation. There is another way: let the network mutate its own structure and keep whatever works. Quantum Neural Horror does that, and lets you watch it happen.Machine learning https://codex-crusader.github.io/writing/browser-hardening/https://codex-crusader.github.io/writing/browser-hardening/Fri, 28 Aug 2026 09:00:00 +0000Bruhswer is a hardened browser for Windows. The most useful thing in it is not a protection. It is the words NOT ENFORCEABLE, printed whenever it cannot actually do what it claims.Security https://codex-crusader.github.io/writing/coding-war-crimes/https://codex-crusader.github.io/writing/coding-war-crimes/Fri, 28 Aug 2026 09:00:00 +0000Most advice about writing good code arrives as a rule you are told to follow. This is the opposite: write the terrible version on purpose, measure exactly what it costs, and let the rule be the conclusion instead of the premise.Software engineering