Appearance

Use Case: Sovereign Edge Sensor Network ​

Deploy dozens or hundreds of HLM-Micro sensor nodes across a facility or critical-infrastructure perimeter. Every node runs on a €5-class microcontroller, every detection is audit-chained, every concept update ships as a small OTA basin patch, and the whole stack runs on EU-sovereign IP without hyperscaler dependency.

Access

Reference firmware, gateway aggregation service, OTA basin-patch pipeline, and partner-specific integration work are delivered through a commercial / partnership engagement. This page describes the deployment pattern and what it earns you operationally.

Who this is for ​

  • Defense / critical-infrastructure program managers needing edge sensing without US cloud dependency
  • Systems integrators building perimeter-monitoring / ISR / acoustic-surveillance products
  • Public-sector and critical-infrastructure buyers in the EU and allied jurisdictions where supply-chain independence is a procurement requirement

The problem you're solving ​

Current edge-sensor networks force one of three unacceptable choices:

ChoiceWhat breaks
Cloud-dependent (AWS / Azure IoT + ML backend)Bandwidth, latency, sovereignty — hostile actor at the cloud layer breaks everything. Not usable in contested RF.
Traditional TinyML per-nodeNo audit trail. No multimodal fusion. OTA update = full reflash. Each node is a single-task siloed classifier.
Expensive SBCs per nodePower + cost per node. Serviceability nightmares. Overkill for what's often a single acoustic classifier.

Contested-environment operation, bandwidth-denied settings, and audit-traceable edge AI are recurring operational requirements for critical-infrastructure and public-sector buyers — and the three force choices above all fail at least one of them.

What a deployment looks like ​

A multi-node sensor network where:

  1. Each node is an ESP32-S3-class MCU running HLM-Micro (€5-class hardware, sub-watt power envelope, sub-MB flash)
  2. Each node classifies locally — no cloud, no backhaul required for core operation
  3. Every classification emits a hash-chain audit trail (~20 bytes/inference) locally
  4. A gateway aggregates classifications with a regional chain of evidence — Merkle-style hash-linked log that a downstream auditor can verify cryptographically
  5. Concept updates (e.g. "start detecting this new drone signature") ship as small OTA basin patches — no full firmware reflash

This is the exact pattern we design around for critical-infrastructure and public-sector pilots.

The stack ​

PieceDelivery
HLM-Micro (pre-trained bases)Model Zoo — Early Access
INT8 quantisation + basin pagingPartnership-level engagement for native ESP32-S3 runtime
Per-node audit digestBundled with commercial release
Gateway aggregation + regional audit logReference service delivered per-deployment
OTA basin-patch pipelineRoadmap v1 — architecturally supported today

How it works ​

1. Pick a vertical-specific base model ​

Pre-trained HLM-Micro checkpoints (Model Zoo) map to common defense-edge verticals:

VerticalBase modelWhy
Acoustic perimeter / counter-UASAudio keyword / acoustic-anomaly baseAudio classification
Industrial monitoringhlm-micro-anomaly-v0Multimodal sensor fusion
Activity / occupancyhlm-micro-har-v0 (real UCI-HAR data)IMU-based classification
RF signature / SIGINT preprocessingPlanned (I/Q sample base)I/Q sample classification

For the acoustic-perimeter scenario below, the base is an audio-classification HLM-Micro.

2. Fine-tune on your signatures of interest ​

If the canonical base doesn't cover your specific signatures (drone models, machinery types, etc.), retrain on your labelled data. Delivery-side tooling takes labelled clips, produces a deployable checkpoint, and signs the weights file. Retraining times are minutes on CPU for small class sets.

3. Deploy to nodes with audit-enabled inference ​

Each node runs the HLM-Micro firmware with your fine-tuned weights. The operational loop on a node is:

  1. Read a sensor window (audio / IMU / ADC).
  2. Run local classification at deploy-mode T.
  3. Compute a ~20-byte audit digest for that inference.
  4. If confidence exceeds the alert threshold, emit (class, confidence, digest) to the gateway.

No cloud involved in the local decision path. The firmware and the C runtime for native on-MCU inference are delivered under the partnership engagement.

4. Gateway aggregation with a regional audit log ​

A gateway device (laptop / RPi / industrial mini-PC) collects events from all nodes and maintains an append-only regional audit log. The log is a Merkle-style hash chain: each new event's regional hash includes the previous regional hash. Any single tampered event breaks the chain forward of that point and is cryptographically detectable on replay.

Operationally this means:

  • A downstream regulator / commander can verify integrity of the entire regional log with a single hash check.
  • An internal review can identify where in the timeline any tamper occurred.
  • Claims of "this detection was made at this node at this timestamp" become binding evidence rather than hand-waving.

5. Replay-verify an incident ​

A claimed detection is challenged ("did node 12 at 03:14 Z actually detect a drone?"). The review:

  1. Retrieve the original sensor window (if preserved — storage policy decision).
  2. Retrieve the node-level audit digest from the regional log.
  3. Retrieve the deployed model weights hash.
  4. Re-run inference on any device with the same weights.
  5. Verify the digest matches → detection is cryptographically confirmed (or not).

This is the same verification primitive as industrial edge and medical imaging, applied at sensor-network scale.

6. OTA concept updates as basin patches ​

A new drone model is observed in-theatre. Rather than full firmware reflash to every node:

  1. Add the new class to your labelled dataset.
  2. Retrain HLM-Micro with the new class.
  3. Diff the resulting weights against the deployed version — identify the specific basins that changed.
  4. Export those basins as a small binary patch.
  5. Push the patch to nodes via standard OTA channels.
  6. Nodes reload the affected basins live — no reboot, no downtime.

The patch size is small enough for LoRa-class long-range radio. This is operationally different from any commodity edge-AI stack today — and it's the main reason the pattern fits contested-environment deployment profiles.

  • On-MCU native inference is partnership-level today. The default pilot shape is host-inference (laptop / gateway runs the model; nodes stream features). Native ESP32-S3 runtime ships under partnership engagement.
  • Certificate mechanism works today in host-inference. Moving it to on-MCU native is an engineering port — straightforward, scoped under the partnership.
  • OTA basin-patch pipeline is architecturally supported but not shipped. The architecture supports it (basins are addressable); the productised patch-build + signed-OTA tooling is on the roadmap.
  • Sovereignty disclosure: Qriton is an EU-incorporated company (Romania). All IP is EU-controlled. Training uses EuroHPC Leonardo BOOSTER (Italian national compute). No US hyperscaler dependency at any layer of the stack.