- Service Definition
- On-device AI with TinyML on Cortex-M, RISC-V & FPGA. Edge AI projects from sensor fusion to computer vision — EU partner network.
Inovasense coordinates Edge AI and sensing projects from requirements to validated hardware and production handover. We help select the processing architecture, integrate sensors and develop the firmware and model pipeline against measurable acceptance criteria.
Define the task before choosing an accelerator
Start with what the product must detect, the acceptable false-alarm and missed-detection rates, response deadlines, energy budget and operating environment. Local inference can reduce dependence on cloud connectivity and limit transmitted data, but latency, accuracy and battery life must be measured on the complete product.
For vibration monitoring, establish representative operating modes and fault data. For computer vision, define lighting, optics, image resolution, defects and production speed. For sensor fusion, verify calibration, timing and the contribution of each sensor to the result.
Hardware and model integration
We can coordinate MCU-based TinyML, application-processor/NPU systems and FPGA acceleration according to the workload. Device selection includes supported model operators, memory capacity and bandwidth, interfaces, software tooling, lifecycle availability and the update path. Datasheet TOPS are a peak capability, not a guaranteed application benchmark.
Quantization and pruning are evaluated against task accuracy and runtime measurements. Language models require memory for weights, runtime buffers and context; a “small” model is not necessarily suitable for a microcontroller. We select an implementation after profiling a representative pipeline rather than promising a fixed accuracy loss or inference time.
Validation and engineering deliverables
The project can include a requirements and architecture document, sensor/PCB integration, embedded software, versioned model builds and reproducible test results. Measure acquisition, preprocessing, inference and output latency, plus complete-board power across relevant operating modes. Record dataset separation and performance under changing environmental conditions.
For safety-related applications, define applicable safety requirements and independent safeguards. An anomaly detector alone is not a validated emergency-stop function. Model and firmware releases need secure update handling, recovery tests and traceability to the validated configuration.
Cloud processing and JustOpen.io
Where fleet monitoring, historical analysis or dashboards are needed, we can coordinate cloud integration and the JustOpen.io IoT data platform. Define which telemetry is sent, how devices authenticate, retention rules and behaviour during connectivity loss. Keep time-critical local tasks independent of unnecessary cloud round trips where the architecture allows it.
Product-specific compliance
Local processing can help minimise data transfers but does not establish GDPR compliance. Assess personal data, retained images, logs and remote support. AI Act scope depends on intended use, risk category and the role of each operator; use the current Commission timeline. Product conformity also depends on the applicable electrical, radio, cybersecurity and sector rules.
We can coordinate the compliance matrix, technical evidence and laboratory work. A component certificate or an accelerator specification does not establish conformity of the final product.
Edge AI validation guide · MCU comparison · CRA checklist · Discuss your sensing project
Primary sources
Technical and regulatory references checked on 1 October 2026.
Guides in this area
Edge AI in Industrial Applications
How to select and validate Edge AI hardware for industrial sensing and vision: latency, energy, model accuracy, update security and EU regulatory scope.
What is Edge Computing? Full Guide
Edge computing architecture for hardware engineers. Processing hardware (FPGAs, GPUs, NPUs), latency tiers, industrial use cases, and EU compliance.
Frequently Asked Questions
What is Edge AI?
Edge AI runs inference close to the sensor or user on a device. It can reduce network transfers and support local operation; latency and accuracy depend on the model and complete hardware/software pipeline.
Can TinyML guarantee years of battery life?
No. Battery life depends on sensors, sampling, inference, radio use, board leakage, battery characteristics and temperature. Validate a complete-device energy budget with measurements.
How is Edge AI evaluated under the AI Act?
By intended use, risk category and the provider/deployer roles. Local execution alone does not establish compliance or high-risk classification.