Immutable model version
Model bytes are content-addressed and signed with their manifest before a rollout can reference them.
Model delivery
Signed OTA rollouts for ML models
A model is not just a file download. Meshanics records its digest and compatibility metadata, stages it into a retained slot, evaluates the candidate locally, and commits or restores the previous model according to the adapter contract.
Good fit
When to evaluate it
- +Computer vision and inference products
- +Teams deploying ONNX or TensorRT artifacts
- +Customers who cannot upload raw inference samples
Product boundary
What the current product does
Slot activation
The Linux model adapter stages a candidate separately, switches the active slot and retains the prior version for recovery.
Local acceptance
A configured evaluation command runs against bounded inputs on the device and returns only its result.
Aggregate comparison
The optional SDK reports count, quality and latency aggregates by version, with minimum sample thresholds and no raw payload transport.
Explicit limit
Meshanics does not train models, manage experiments, label data, or inspect inference samples. The application remains responsible for producing meaningful health and quality signals.
Console proof
See the operational record
Questions
Does Meshanics require changes to application code?
No. The static Linux agent installs on the device and activates signed payloads through typed adapters. Health behavior and recovery limits are configured for the artifact, not embedded in customer application code.
Does Meshanics certify CRA compliance?
No. Meshanics records and exports evidence for the manufacturer's compliance work. The manufacturer remains responsible for product scope, risk decisions, conformity assessment, reporting, and declarations.