Two systems, both measured.

One is finished enough to try in a browser. The other is research that has not yet beaten its own baseline, and says so.

Prototype · public demo · Apache-2.0

ayaka-hand

A prosthesis control system that pre-shapes the grip before contact. A wrist camera looks at the object being reached for; the muscle signal is reduced to a single binary trigger. The user decides when to close, the vision system decides how.

The student model gets nothing but pixels — no object detector telling it what it is looking at — and still beats a baseline that is handed the correct object name.

Read the write-up →

Measured

Grip accuracy: 72.3 %
Force accuracy: 81.1 %
Parameters: 2,232,839
Model size: 8.9 MB ONNX
Latency budget: ≤ 125 ms

Active research

Hime

A translation pipeline for long-form literary Japanese. Instead of one large model in a data centre, several small fine-tuned models handle each paragraph in sequence: draft, merge, polish, review. It runs on a single consumer GPU, fully offline.

The fine-tuned models do not yet reliably beat their own base model. The most recent supervised fine-tune measured as a regression. What does work is the evaluation infrastructure — and it once caught a positive result of its own that had to be withdrawn.

Read the write-up →

Setup

Four-stage pipeline
Single consumer GPU
Fully offline, EPUB in and out
De-leaked judge harness