Edge and GPU inference
AI at the edge.
Open-weight models on GPU compute we operate, including NVIDIA Jetson devices at the edge.
Hardware
Open-weight models, run on hardware we operate.
Our own work runs on NVIDIA CUDA GPUs we operate, scheduled by a job router we wrote. Before a machine gets work, it passes a real computation.
Deployment
Edge inference on NVIDIA Jetson.
Where each component runs, and which model calls leave your environment, is written down in a data-boundary document before anything is built. Edge hardware is one of the options.
Pipeline
A ledger you could reproduce.
Jobs are split into checkpointed units, and a failed unit is retried on a different machine.
- Status
- Runs our own work.
- Boundary
- GPU compute we operate.
- Acceptance
- No machine gets work until it passes a real computation. Every job leaves a ledger.
The evaluator
We taught a smaller model to make the same calls as the evaluator.
The evaluator is Jev, from TypeSafe AI. The smaller model is ModernBERT-large, distilled from 68,828 Jev-labeled training examples, so the same calls run on compute we operate.
It is measured as teacher agreement, not accuracy, on two cohorts, with zero-shot Laya as the baseline.
In-KB n = 4,387
Unseen KB n = 1,301
The limit. Rare flags are not claimed.
- Status
- Measured. The figures are published on OptiVis AI Labs.
- Boundary
- Trained and run on GPU compute we operate.
- Acceptance
- Teacher agreement on two cohorts, against a zero-shot baseline.
Engagement
How an engagement is priced.
Diagnostic
$7,500
Ten business days, one workflow. Credited in full against a build signed within 60 days.
Production build
From $25,000
Fixed price, accepted against tests written before the build.
Operate
From $2,500 a month
Bounded and optional.
Start with the diagnostic.
One workflow, with its data boundary written down before anything is built.