Projects

Engineering projects with technical deep-dives by Yago Mendoza.

- Class notes in, Anki cards out: Raw class notes go into a folder, and an LLM agent turns them into a structured dictionary and its Anki cards under a written specification. A Python script decides which cards each word requires and refuses any batch that misses one; the deck reaches Anki desktop through AnkiConnect. Each batch returns a review to the learner, and his answers to it start the next one. The system is MIT and the accumulated content CC BY. - How InfraPhysics is built: a site with no server, organised around its wiki: InfraPhysics is a personal site built as one repository and one build: about 2,800 markdown files, a wiki of more than 1,500 notes among them, become static files served with a thin layer of edge functions. This write-up follows one paragraph from the keyboard to its five kinds of reader, then reconstructs from 1,148 commits the decisions that shaped each layer and what each one cost. It began in March as a diary of the first eighteen days and was rewritten in October. - Forecasting visual auras from event streams: a simulation-first study: A simulation-first rehearsal for forecasting a visual epileptic aura from the small visual events that may precede it, and a stricter measurement of what a model would actually be allowed to claim. - A CRM for a health-tech startup, built solo in 72 hours: A bespoke CRM for a health-tech startup, designed and shipped solo in 72 hours as an experiment in AI-assisted engineering. Its 23 tables are held together by 8 explicit state machines, and a stale-detection clock flags the relationships going cold. - An investment dashboard with no build step: vanilla JS and Chart.js: A vanilla JS investment dashboard built in one session with Claude Opus 4.6, with no build step: the whole app is a folder you double-click. Its second half rebuilds four years of cash flows from purchase records alone, framed as a Material Requirements Planning problem and validated to the cent. - Reproducing a Stable Diffusion model that invents kanji for English words: An attempt to reproduce enpitsu's 2023 Stable Diffusion fine-tune, whose invented kanji for English words hardmaru shared that January, by fine-tuning Stable Diffusion 1.5 on rented GPUs for $66 in all. For most of the runs my model learned the look of kanji and ignored the caption, until a 256-character overfitting test showed it was short of passes per caption; after that the meaning showed in one radical (氵 for Water, 金 for Gold). Training on enpitsu's full dictionary of 10,358 kanji brought his kind of pun (竹 for Nicholas Cage, from 籠, cage) without his drawings, and a blind judge preferred the glyphs of his model 17 to 8. - TokenWalk: a step-by-step simulator of local LLM inference: The console behind the transformers article. Set up a run, press Play and watch the request go from the disk to the last token; click anything to see what it is. It also tells you which models fit on which machines, and can run the same request on your own Ollama. - TrialGPT: from a medical report to a shortlist of eligible trials: A clinical-trial matching tool built at Foundation 29 for patients with rare and hard-to-diagnose conditions. TrialGPT reads a case or an uploaded report and checks it against the inclusion and exclusion criteria of trials in ClinicalTrials.gov and the European CTIS register, one clause at a time, keeping a third verdict for what the report does not say. The write-up follows the 2025 build through its failures, then reads how the live product works a year on.