Sovereign ONNX Edge Mesh
On-Device Inference Without Cloud Egress
Cloud-dependent AI inference creates latency, network dependency, and data sovereignty risks for autonomous drone operations in contested or bandwidth-limited environments. The SOEIM engine (M220) deploys ONNX Runtime models to edge nodes via a KSL-signed distribution pipeline, executes inference locally, and aggregates results through a sovereign mesh that never routes raw sensor data off the originating node.
Capability specification
- 01
KSL-signed ONNX model distribution to edge nodes with SHA-256 model fingerprint verification
- 02
Local ONNX Runtime inference: raw imagery and sensor data never leave the edge node
- 03
Mesh result aggregation using hashed embeddings only (KURAL 23)
- 04
Model version lifecycle management with revocation and rollback
- 05
54-test green suite covering model fingerprint, inference boundary, and mesh aggregation
How it works
Model Distribution
ONNX models are signed with a KSL key and distributed to registered edge nodes. Each node verifies the SHA-256 model fingerprint before loading. Unsigned or fingerprint-mismatched models are rejected.
Edge Inference
Inference runs entirely on the edge node. Raw sensor data — imagery, GPS, acoustic recordings — never leaves the node. Only structured inference outputs (EPPO code, confidence, bounding box dimensions) are emitted.
Mesh Aggregation
Structured inference outputs from multiple nodes are aggregated via hashed embeddings and broad region prefixes. The mesh coordinator produces consensus records without raw data cross-contamination.
Standards we follow
- STD-01
ONNX Runtime open standard (inference engine)
- STD-02
NIST FIPS 180-4 — Secure Hash Standard (model fingerprint SHA-256)
Areas served
This capability is deployed across 14 operational regions. Regulatory alignment details vary by jurisdiction — consult engineering for jurisdiction-specific deployment guidance.
Frequently asked questions
What happens to inference if the edge node loses network connectivity?
Inference continues locally without interruption. The edge node queues structured output records for mesh synchronisation when connectivity resumes. The platform's offline-first design means network loss affects aggregation latency, not inference availability.
Talk to engineering
For capability evaluation, integration guidance, and deployment scoping, submit a brief to the engineering team.
Submit engineering brief