13 kV RMU Fault Detection
On-device DSP + classifier for partial discharge - evidence-rich alerts, sub-second decisions.
Read case study
Capability
Edge AI Development in Australia
Inference on constrained hardware against real latency and power budgets - quantised models, measured on the target, not benchmarked in a notebook. When the constraint is milliseconds or milliwatts, cloud inference is not a fallback plan - and we build the alternative.
Scope
What edge AI development covers - and what it doesn't.
Edge AI is inference that runs on the device that produces the data - because the round trip to the cloud is too slow, too expensive, or too unreliable. We treat it as a systems problem: the model, the runtime, the hardware and the sensor are one design decision.
In scope
- On-device inference on MCU-class hardware (TFLite Micro, ExecuTorch)
- Jetson-class embedded GPUs (Orin, Xavier, Nano)
- Coral Edge TPU, Hailo, Kneron and other accelerator ICs
- Model quantisation (INT8, INT4), pruning and structured optimisation
- Deployment tooling: ONNX Runtime, TensorRT, TFLite, ExecuTorch
- Sensor + model co-design (camera choice, framing, illumination decisions)
- On-device continuous learning and safe update patterns
Honestly out of scope
- Training foundation models from scratch (we fine-tune, we adapt, we do not pre-train)
- Pure cloud inference or hosted-only ML-ops platforms
- Managed 24/7 ML monitoring operations after handover
- Research-only work with no deployment target
Outcomes
What you get out of it.
Edge AI succeeds or fails on measured behaviour on the target, not benchmark scores on a workstation. What our edge AI programs consistently deliver:
Latency budgets that hold at rate
Models sized to their real deadline. Our weld-inspection system runs inference in 38 ms inside a 45 ms window at line speed, every cycle. No frame skips, no fallbacks.
Power budgets that hold in the field
Quantised inference paths designed for coin-cell and solar-powered nodes. Duty cycle chosen so the model runs when it needs to, sleeps when it does not.
Accuracy that matches the deployment, not the paper
Models validated on data collected under the deployment conditions (illumination, background, occlusion), not on curated benchmark sets. Failure modes documented before deployment.
Update paths that survive the field
Signed model updates over OTA with rollback on inference-quality regression. A bad model does not brick the fleet.
Process
How a edge AI program runs here.
Four phases with a real deliverable at each gate - you always know what you paid for and what ships next.
PHASE 01
Discover (paid week)
We start from the constraint that binds - power, latency, thermal, certification - and design backwards from it. You leave with a written architecture, a budget range and the risks named, whether or not we build it.
PHASE 02
Design
Schematics, mechanical and firmware architecture proceed in parallel. High-risk blocks get simulated or breadboarded before the full layout commits.
PHASE 03
Build
Iterative revisions against real bench and field testing. You see every revision, not just the last one. Integration is continuous, not a phase.
PHASE 04
Deploy
Pilot in the field, closure with the contract manufacturer, production test procedures, and a commissioning-grade handover pack.
Technologies
What we build on - chosen per constraint, not per preference.
The platforms we reach for most. If a project needs something not on this list, we say so - the tool is chosen for the constraint, never to fit our habits.
TensorFlow Lite / TFLite MicroONNX RuntimeTensorRTExecuTorchNVIDIA Jetson (Orin, Xavier, Nano)Coral Edge TPUHailo-8KneronSTM32 Cube.AIESP32-S3 AIPyTorch quantisation (INT8/INT4)MLflowDVCW&B for experiment tracking
Industries
Where edge AI we ship runs today.
AgriTech
Crop-vision models on outdoor devices, disease detection at rate.
Energy
Fault-classification models on grid instrumentation with sub-second decisions.
Robotics
Perception stacks on autonomous platforms with hard safety constraints.
Smart manufacturing
Line-side inspection with deterministic cycle budgets.
Deliverables & IP
What ships to you at handover.
Every edge AI program hands over: trained model weights (target-native format), quantisation and calibration data, inference wrapper code with API documentation, deployment runtime with example integration, model card documenting training data, accuracy and known failure modes, and a signed update pipeline for future model refreshes. All foreground IP transfers on payment.
Case studies
Programs we shipped in this space.
Every entry links to the full case study - constraints, what we built, and what it measured afterwards.
Inline Weld Seam Inspection
Vision inference in 38 ms inside a 45 ms line window; 100% inspection coverage.
Read case study
Autonomous Zero-Turn Mower
Perception + planning on Jetson-class hardware with 360° safety coverage.
Read case study
Compliance
AI transparency and safety by design.
Every edge model we ship comes with a documented model card: training data provenance, accuracy on held-out validation, known failure modes, and boundaries of intended use. For safety-adjacent deployments (mobility, medical, industrial control) we design the fallback path - what happens when confidence drops below threshold - explicitly, not implicitly.
Australian consumer AI is increasingly under scrutiny for privacy and transparency (Privacy Act reforms, sector guidance for health and financial services). We build with data-minimisation and on-device processing as the default - most of what we ship never sends raw data to a cloud - and document the data flows for buyer review. Where a model must meet a specific regulated-industry expectation, we align architecture and validation to it.
FAQ
Edge AI Development in Australia - straight answers.
Model adaptation and deployment on existing hardware: AUD $10,000-$30,000. Full edge AI product (device + model + update pipeline): AUD $40,000-$150,000+. See our <a href="/blog/ai-development-cost-australia.html">AI development cost guide</a> for detailed bands.
We fine-tune and adapt where a foundation model or open-weights base exists (which is most of the time). Full from-scratch training is rare in production edge AI and usually the wrong economic choice. We say when it is not.
Constraint decides. Vision at 30 fps with modest power - Jetson class. Ultra-low-power keyword or anomaly detection - MCU with a NN accelerator. Vision at low fps in tight thermal envelopes - Coral or Hailo. Every choice is made against the actual latency and power budget.
On data collected in the deployment environment - not the curated benchmark set. We build validation datasets that reflect the illumination, occlusion, and edge cases the model will see, and document accuracy against them.
You do. Model weights, training data (where legally and contractually possible), quantisation configuration, and inference code transfer on payment. Pre-existing foundation-model weights we build on follow the base model’s licence.
Yes - we design the model update pipeline to accept new data with human labelling, retrain, and push signed updates to the fleet. The retraining cadence is your call; the plumbing is built once.
Why Incendio
One team, no seam between vendors.
Edge AI detached from the hardware that runs it produces models that pass on the workstation and miss deadlines on the target. Because we lay out the boards and write the firmware ourselves, model choices are made against the real inference silicon and the real power budget - and the numbers we quote are the numbers that ship. Judge the build, not the pitch: open our live demos.
Related practices: embedded systems development, computer vision, robotics, IoT development.
Start
Tell us the constraint that worries you most.
A latency budget, a power budget, a certification date. We reply within one business day - and we’ll say so if we’re not the right team.