lab
Research topics and open source.
What we're investigating, and the projects we intend to share. Listings grow as work leaves the lab.
Open source
4 projects
- active
virtio-accel
A virtual accelerator protocol that ships with the machinery to run it: guest, device, transport, queue, and TOSA 1.0 layers in no_std Rust, plus host backends that put the same artifact on Apple's Neural Engine, Intel NPUs, and Qualcomm Hexagon. Fifteen crates, protocol 1.0 frozen.
- active
autocommit
Local AI commit and PR drafting — analyzes git diffs on-device and writes structured messages through an interactive terminal workflow.
- active
unsigned-float
Unsigned floating-point types for non-negative domains — compact IEEE-like layouts without a sign bit.
- planned
reference agents
Small, auditable agents that demonstrate scoped grants — refactor, docs, vision — without ambient authority.
Research topics
3 threads
- active
Cross-device tensor parallelism
Shard attention and FFN matmuls across a user’s own devices so they compute the same layer together — pooling FLOPs and weight memory past what any single phone or laptop can hold. Edge TP is communication-bound and device-skewed; the open question is when the aggregate still beats running alone, and how to place shards across a heterogeneous personal fleet.
- active
Forward-pass model adaptation
Zeroth-order optimizers estimate gradients from loss differences under parameter noise, so fine-tuning needs only forward passes and inference-level memory — no backprop activation store. LOZO pushes this further with low-rank perturbation structure; we’re using that family of methods so models can keep improving on-device as they are used.
- active
Anticipatory personal surfaces
First, learn behavioral patterns well enough to predict the next want. Second, run an agent loop that builds the UI, aggregates the data, and implements the capability before that moment arrives — so when the user reaches for action A, the feature for A is already live.