Lab
Demos of the search and retrieval actually running on this site, and below them the side projects that never grew into case studies. Expect rough edges.
Running here
Semantic search demo
Try a query. This is hybrid retrieval over my essays: results are fused from vector similarity and keyword search, so a post can rank on meaning, on wording, or both. Each result says which one caught it.
Toy text similarity
A tiny cosine-similarity demo, scored on word frequency rather than on embeddings. Reword one of these to mean the same thing without reusing its words and the score collapses — that gap is the whole reason retrieval uses embeddings instead of this.
Elsewhere
Quantum sensors, numerology, a Fediverse server, and a deck of cards for when the work stalls. None of these are a portfolio piece. All of them run.
- Creative block
Grasping Straws
A deck of lateral-thinking prompts in the spirit of Brian Eno and Peter Schmidt's Oblique Strategies, with entirely original card text. Tap, get a card, tap again. Static Astro, and the built output ships zero framework JavaScript. - Esoterica
Numen
Numerology read across both the Pythagorean and Chaldean systems, with tarot birth cards and astrological correspondences, styled as a dark grimoire. Nonsense taken seriously enough to implement correctly. - Physics
qsensor-sim
Simulated quantum inertial sensors for GPS-denied navigation research — trajectories with decoherence effects, run through Kalman filtering. Built because the hardware is scarce and the signal processing downstream of it should not have to wait. - Film
Movie Conceptualizer
Fountain screenplay in, shot list and storyboard out, via a set of agents that each handle one stage of pre-production rather than one model doing all of it badly. - Music
Album Conceptualizer
A workspace for building coherent concept albums — structure, theory, and export handoff. Deliberately not a one-prompt song generator; it is strongest at the album-level coherence that those tools ignore. - Music
Free VST Plugins
A curated, containerized catalogue of free audio plugins, because the existing lists are affiliate-farmed and half the links are dead. - Books
Kindle Classics
Scripts that pull the most-downloaded public-domain classics from Project Gutenberg via Gutendex, convert them, and sideload a whole library onto a Kindle without going through a store. - ML
HuggingFace portfolio
Sixteen published projects on the Hub — datasets, fine-tuned models, and interactive Gradio Spaces across NLP, vision, and generative work. - Music
SpotifyForge
A CLI and REST API for people who curate playlists seriously, backed by a local cache database and a scheduler so the automation survives Spotify's rate limits. - Records
discogs
A Python library and CLI for the Discogs API that syncs a collection and wantlist into a local cache — the groundwork for recommendations that know what is already on the shelf. - Discipline
75 Create
The 75 Hard format bent toward creative work: five daily tasks, seventy-five days, and a log of the output at the end. Runs entirely in the browser with no backend and no account. - Fediverse
polyvox
A Fediverse server MVP with timelines, moderation workflows, privacy export and delete, and NodeInfo/WebFinger discovery. In progress, and mostly an excuse to read the ActivityPub spec properly.