Research · 10 projects
Things I measured
Experiments with language models, agents and wearables. Each page explains how the thing works, what I measured, and where it stops, with the code to check it. None of these are peer-reviewed papers.
01
Language models
Making open models do more on less: one forward pass instead of generation, sparse experts on one desk machine, small fine-tunes that run on every keystroke.
- Sep–Oct 2026Open source82/111hard items, one pass
From Transformers to Decision Models
Turning an open language model into a classifier that answers from a fixed set of options in one forward pass, and measuring what that shortcut costs.
How it works - Apr 2026Open source6,874tokens/s with fused dispatch
Small Models, Sparse Experts
Pre-training a 4.73B-parameter sparse mixture-of-experts model on one Mac Studio.
How it works - Jul 2026Open model13/16course-corrections, up from 10
On-Device Language Model Fine-Tuning
A 350M-parameter dictation clean-up model that keeps the correction when you change your mind mid-sentence.
How it works - Jul 2026Open source11×faster than the CPU path
Brain-Response Models on Apple Silicon
Running Meta's TRIBE v2 brain-encoding model quickly on a Mac by porting its video encoder to MLX.
How it works
02
Agents
Testing agents the way they fail in practice: on a real phone line, and against the documents they trust.
- May 2026Experiment~600 msto first audio per turn
Self-Learning Voice Agents
Two local voice agents call each other over Telegram, a model grades the call, and a loop tunes the pipeline.
How it works - Jul 2026Public code35/38planted lies found by the best agent
Can Agents Spot Planted Lies?
A benchmark that plants false claims in the docs, config and source of a real open-source monorepo, asks coding agents to audit it, and counts how many of the planted lies they find.
How it works
03
Body, biology and learning
Work away from the keyboard: a wearable that bends the elbow, proteins as logic, and how medical students learn hard ideas.
- 2024Prototype2.26×pull from four strands, after friction
Discreet Artificial Muscle
A wearable prototype for low-profile muscle assistance: a shoulder-mounted stepper winds fishing line through sewn-on hooks to bend the elbow.
How it works - 2026PausedShared switcha computational architecture and its limits
Computational Protein Logic Gates
A shared-switch protein XOR model: how its readout differs from existing gates, and where sequestration fails under input imbalance and affinity uncertainty.
Computational working paper. Physical readout unresolved; no protein gate built or tested.
How it works - Oct 2026Open source1.5 s52-nt RNA on an M4 GPU
RhoFold+ on Apple Silicon
How to run RhoFold+ RNA 3D structure prediction on an M-series Mac: PyTorch on the Apple GPU (MPS), one setup script, a 52-nt sample in seconds. MLX path experimental.
How it works - 2024Poster
Learning by Making: AACA 2024 poster
A qualitative study of how medical students learn difficult topics, and the role of creative models in active learning.
Third of five authors on the abstract in the AACA 2024 programme.
Read the post