Apollo —
Multi-Branch Temporal Attention Network
Apollo is the flagship AutomataNexus diagnostic appliance — a purpose-built edge inference device for commercial HVAC predictive maintenance. It ingests live BACnet/Modbus sensor telemetry from building automation systems, processes 57-dimensional temporal sequences through a multi-branch architecture with parallel dilated convolution streams at 3 time scales, cross-branch attention fusion, and a 3-head output predicting imminent faults, degradation trends, and optimal setpoint corrections. Deployed across 40+ commercial sites in Indiana and Ohio.
Abstract
Background
Apollo is the flagship AutomataNexus diagnostic appliance — a purpose-built edge inference device for commercial HVAC predictive maintenance. It ingests live BACnet/Modbus sensor telemetry from building automation systems, processes 57-dimensional temporal sequences through a multi-branch architecture with parallel dilated convolution streams at 3 time scales, cross-branch attention fusion, and a 3-head output predicting imminent faults, degradation trends, and optimal setpoint corrections. Deployed across 40+ commercial sites in Indiana and Ohio.
Approach
The model was trained in AxonML (a pure-Rust deep learning framework) and compiled through the Hailo Dataflow Compiler (DFC 3.33.1) targeting Hailo-8 silicon. Post-training INT8 quantization was applied during the DFC compilation pass with production telemetry calibration data. The resulting Hailo Executable Format (HEF) binary executes on Hailo’s fixed-function dataflow architecture with deterministic latency and zero framework overhead at the edge.
Results
On production hardware (Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289)), Apollo achieves 1,475 FPS (hw_only) with 0.706 ms hardware latency at 2.12 W average power draw.
Conclusion
Apollo is production-ready as a single HEF binary deployed to edge devices with no external dependencies beyond the HailoRT vendor runtime. The model meets real-time latency requirements for its target hvac predictive diagnostics application.
| Model | Apollo |
| Domain | HVAC Predictive Diagnostics |
| Architecture | Multi-Branch Temporal Attention Network |
| Target silicon | Hailo-8 |
| Measured on | Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289) |
| DFC compiler | 3.33.1 |
| Framework | AxonML v0.6 (pure-Rust, CUDA + CPU backends) |
| Author | Andrew Jewell Sr. · ORCID 0009-0005-2158-7060 |
| Organization | AutomataNexus LLC · Fort Wayne, Indiana |
Executive overview
Apollo is the flagship AutomataNexus diagnostic appliance — a purpose-built edge inference device for commercial HVAC predictive maintenance. It ingests live BACnet/Modbus sensor telemetry from building automation systems, processes 57-dimensional temporal sequences through a multi-branch architecture with parallel dilated convolution streams at 3 time scales, cross-branch attention fusion, and a 3-head output predicting imminent faults, degradation trends, and optimal setpoint corrections. Deployed across 40+ commercial sites in Indiana and Ohio.
Network I/O
Input: 57-channel sensor vector (1 timestep, real-time BACnet/Modbus poll). Output: 3-head predictions (fault class logits + trend score + setpoint deltas).
Architecture
Multi-Branch Temporal Attention Network
The Apollo architecture processes 57 sensor channels (supply/return temperatures, pressures, flows, valve positions, damper angles, electrical current, vibration) through three parallel temporal branches operating at 1x, 4x, and 16x dilation rates. Each branch contains 4 depthwise-separable convolution layers with batch normalization and ReLU gating. Branch outputs are fused via a learned cross-attention mechanism (implemented as 1x1 convolution over concatenated features), followed by a global average pooling and three independent classification heads: (1) imminent fault detection (8 fault classes), (2) degradation trend scoring (continuous 0-1), and (3) setpoint optimization (delta values for supply air, chilled water, hot water). The model is quantized to INT8 with DFC-calibrated ranges from 10,000+ hours of production telemetry.
Compilation constraints
All AxonML models targeting Hailo silicon are compiled under the fixed-function dataflow constraints: no dynamic control flow, no variable-length dimensions, all activations representable in INT8 after calibration, and no operations requiring dedicated softmax hardware (replaced with ReLU gating or depthwise convolution equivalents where necessary).
Silicon performance
Measured on production hardware via hailortcli benchmark with 5-second sustained inference. Device: Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289).
| Metric | Measured Value |
|---|---|
| FPS (hw_only) | 1,475.01 |
| FPS (streaming) | 1,475.00 |
| HW Latency | 0.706000 ms |
| Power (streaming avg) | 2.11500 W |
| Power (streaming max) | 2.12700 W |
| Power (idle) | 0.74979 W |
| Quantization | INT8 (post-training, DFC calibration) |
| DFC Compiler | 3.33.1 |
| HailoRT | 4.20.0 |
| Measured On | Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289) |
Deployment
Deployed as a single HEF binary. No ONNX runtime, TensorFlow Lite, or Python inference stack required at the edge.
| Target silicon | Hailo-8 |
| Measured on | Hailo-8 M.2 (P/N: HM218B1C2FAE, S/N: HLDDM2A234600289) |
| DFC compiler | 3.33.1 |
| Quantization | INT8 (post-training, production telemetry calibration) |
| Runtime | HailoRT (vendor runtime) |
| Edge platform | Raspberry Pi 5 + Hailo AI HAT+ (M.2 Key M) |
Deployment procedure
Copy the .hef binary to the target device. hailortcli run loads the HEF directly into the Hailo-8 dataflow engine over PCIe. Inference begins immediately with deterministic per-frame latency. No model conversion, graph optimization, or warmup phase required.
References
- Jewell, A. (2026). AxonML: A Pure-Rust Deep Learning Framework for Edge Inference. AutomataNexus LLC. Technical whitepaper.
- Hailo Technologies Ltd. (2024). Hailo Dataflow Compiler User Guide. DFC v3.33.1.
- Hailo Technologies Ltd. (2024). Hailo-8 Product Datasheet.
AutomataNexus LLC · Fort Wayne, Indiana · andrew.jewellsr@automatanexus.com
Andrew Jewell Sr. · ORCID 0009-0005-2158-7060
May 2026 · All rights reserved.