Musical analysis — Libretto
Drop a MIDI file, get it scored across 29 weighted axes: form, harmony, melody, rhythm, texture, coherence. The analysis runs in your browser, the file goes nowhere.
Try it ↗Laboratory
The lab is open. So are the experiments.
Three test benches and a dashboard. Two of the demos execute in your browser, not on my server: the file you drop goes nowhere. It is the principle this site argues for, applied to its own shop window.
The gateway, in motion.
Drive the simulation: send requests, reserve RAM for audio, watch LRU eviction decide which model to unload.
Trigger a request
What the simulation shows: three models share an 80 GB budget. The conversational model is pinned — it is never unloaded. When a request needs an absent model and there is no room, LRU eviction unloads the least recently used. A 24 GB audio reservation forces several unloads at once.
Calibration — tabgate
Can your classifier's probabilities be trusted? ECE, reliability curve and gate verdict, with scikit-learn running in WebAssembly.
Try it ↗The measurements
The ecosystem's public dashboard: every figure with its measurement date and the command that produced it. No number is typed by hand.
See the measurements →Your data cannot leave the building?
That is precisely the problem I solve. A 30-minute call is enough to scope an audit.