llm-router
Built · publication pending
Route calls. Keep the learning.
Choose a model by cost and quality. Retain consented traffic for training and evaluation.
Independent AI infrastructure
Turn the intelligence you rent into capability you own.
Tools for running, training, and controlling AI on hardware you own.
Explore the work
Tools shipping. The full loop is in development.
Distributed compute
MLX meshMLX mesh / research
Train across Macs and CUDA devices.
Building distributed inference and training across Apple silicon and NVIDIA GPUs. Place work on the hardware you control, exchange model state, and coordinate updates within one cohort.
Macs + CUDA · peer-to-peer transport · operator-owned model state.
Explore the mesh ↗Conceptual topology
Selected work
05 projects
cove
v0.6.0 · shipped
A clean Mac for each run.
Fork a macOS guest, run an agent or CI job, keep the artifacts, and discard the child.
mlx-go
Built · publication pending
Inference and training in Go.
A Go MLX runtime for inference and training, starting with Apple silicon.
lib/nvidia
In development · source private
Use NVIDIA GPUs directly from Go.
Compile CUDA kernels, call math libraries, and collect GPU profiles without cgo.
ModelIR
Experimental · source private
Bring model definitions into Go.
Inspect supported model definitions, compare their structure, and generate Go implementations.
The approach
The model, the data, and the decision to leave should be yours.
We’re building a path from rented model calls to a model you can keep. Route requests, retain consented examples, then train and evaluate on hardware you control.
The test is practical: disconnect from the frontier APIs and measure how much of a real workload your own model can still do. The complete loop is in development.
How we measure ownership ↗