NexusEdge Hailo Edition (Community) — on-device building-automation control on Hailo-8
We’re releasing the Community edition of NexusEdge Hailo Edition (NEHE) — the same Rust controller featured in our case study (linked below), with a few changes for the community and a free tier so anyone with a Hailo-8 and a Raspberry Pi can run it. The Hailo accelerator is central to how it works, so this seemed like the right place to share the architecture.
What it is
NEHE is a Rust edge controller that replaces a traditional DDC/BAS controller on a Raspberry Pi 5 or CM5 with a Hailo-8/8L (and Hailo-10H where available). It reads physical I/O — temperatures, pressures, currents, relays, analog outputs — runs closed-loop control, and uses the NPU for the inference half of the workload: fault detection and diagnostics, setpoint optimization, and anomaly detection, running locally on the building’s own hardware rather than in the cloud.
The design goal was a single static binary with no Python runtime, no GC pauses in the control path, and on-device inference so the controller keeps running with full intelligence even when its network does not. This is the same controller from the case study; the Community edition trims it to a free tier suitable for evaluation and self-hosting, with the managed/Pro features available separately for those who want them.
Where the Hailo NPU fits
- Inference at the edge. Diagnostic and optimization models run on the Hailo-8 alongside the control loop. They’re small and latency-bound rather than throughput-bound; the NPU keeps them off the CPU so the deterministic control path stays clean.
- A model hub on the device. Models are versioned and loadable at runtime, so a controller’s behavior can be updated without reflashing firmware.
- Live silicon telemetry. The console streams real NPU utilization, temperature, and inference timing from the device — useful for confirming a model behaves on real hardware the way it did on the bench.
Platform and architecture
- Targets: Raspberry Pi 5 and CM5 (Bookworm, aarch64) with Hailo-8/8L; Raspberry Pi 4 (Bullseye, armhf) for the headless control role.
- Control engine: a trait-based engine with real control laws — staged AHU/RTU, lead/lag pumps, boilers and chillers, VAV, DOAS, heat pumps, and natatorium control among them — configured declaratively per site.
- Resilience: input filtering with last-good hold through sensor faults, a fail-safe state machine, and a store-and-forward buffer so telemetry is never silently lost.
- Integration: BACnet/IP and Modbus for existing equipment.
The console
The management console provides multi-tenant, role-based fleet management: live metrics, analytics in real engineering units, equipment control and setpoint changes, signed control-logic deployment, one-click commissioning of a fresh device, and an on-device LLM-assisted diagnostics view. It also surfaces the NPU telemetry and model-management pages described above. The free tier covers single-controller use; multi-node and managed features are part of the paid tiers.
Case study
The full Hailo case study on the deployment:
Availability and verification
Public releases are published at:
Builds are produced in CI and ship with SLSA Build Level 3 provenance and signed checksums, so a downloaded binary can be cryptographically traced to the exact source revision and workflow that produced it.
Notes
This is a commercial platform with a free Community tier rather than an open-source library, but the Hailo integration patterns — runtime model loading, streaming NPU telemetry, and keeping inference off the real-time control path — are the kind of thing this forum discusses, and we’re glad to answer questions about how we approached them on the Hailo-8.
For anyone interested in the model side, our Rust-native ML framework AxonML — which we use to build and convert the lightweight edge models that run on the Hailo-8 — is published separately at:
The AxonML + Hailo whitepaper covers the model build/convert pipeline and the design decisions behind it:
https://axonml-hailo.pages.dev/AxonML-Hailo-Whitepaper#cover
— AutomataNexus