Edge AI
Aplark Edge
Run intelligence where the mission happens.
Aplark Edge is intended to close the gap between the model that was assured and the model that actually runs on the target — a different artefact, at a different numeric precision, under a thermal envelope the laboratory did not have.
The problem
The assured model is not the deployed model
Evidence is generated against a model in a laboratory. What deploys is a quantised, compiled, target-specific artefact running on hardware that throttles. Those are not the same system, and the difference is not a rounding error: quantisation changes decision boundaries, and thermal throttling changes latency distributions in the tail where the timing requirements live.
An assurance argument that does not address this has a defeater sitting in plain view, and a reviewer will find it.
How it works
How it is intended to work
The first piece — verifying that the artefact digest at the deployment boundary is the digest that was assessed — requires no special hardware and is being pulled forward into Aplark Assure rather than held back for a separate product.
The measurement work is what needs the hardware: characterising the disagreement set between full-precision and quantised inference, and profiling latency under sustained thermal load on representative targets. That is being scheduled against real hardware availability, not against a marketing calendar.
Capabilities
What it does
Artefact identity at the boundary
Hash verification refusing any artefact whose digest does not match the assessed one, every time, offline. Being brought forward into Aplark Assure.
Quantisation disagreement sets
Intended to enumerate the inputs where reduced-precision inference disagrees with the assured baseline, so the delta is a measured set rather than an assumption.
Thermal and latency profiling
Planned characterisation of tail latency under sustained load on representative targets, since the timing requirement lives in the tail and the tail is where throttling shows up.
Who it is for
- Programmes deploying to constrained or embedded targets
- Teams whose timing requirements are stated at high percentiles