AI & IoT Platform Integration for Connected Hardware
Edge AI on devices, cloud ML pipelines, and integration with AWS IoT, Azure, ThingsBoard, Grafana and more
AI and IoT platform integration connects your devices to the platform that stores, shows and acts on their data, and adds machine learning where it helps. Electuni integrates hardware with AWS IoT, Azure, ThingsBoard, Grafana, Firebase and similar platforms, and builds edge AI, cloud ML and dashboard AI features for product teams.
Design → Prototype → Production
Key Facts: AI & IoT Platform Integration
- Edge AI
- TFLite Micro, ESP-DL, Cube.AI
- Cloud
- AWS IoT, Azure IoT
- Dashboards
- ThingsBoard, Grafana, Blynk
- Data
- TimescaleDB, InfluxDB
- App Backends
- Firebase, Supabase
- AI Features
- Anomaly, Forecast, LLM
What You Get
- Device-to-platform integration: provisioning, secure MQTT or HTTPS connectivity and the data model on the chosen platform
- Dashboards and alert rules on the platform, or a custom dashboard when the product needs its own interface
- Edge AI models converted and deployed to the device, with the firmware that runs them and reports results
- Cloud ML pipeline for training and retraining on collected device data
- AI features in the dashboard, such as anomaly alerts, forecasts or LLM-based summaries of device data
- Documentation of the data flow, model versions and platform configuration for your team
How We Approach AI & IoT Platform Integration
Platform integration starts with the data: what each device sends, how often, and what the people using it need to see or be alerted to. That decides whether an off-the-shelf platform such as ThingsBoard, Blynk or Ubidots is enough, whether a cloud service such as AWS IoT or Azure IoT fits your existing infrastructure, or whether a self-hosted stack with TimescaleDB or InfluxDB and Grafana gives you more control over cost and data.
AI is added where it changes what the product can do. On the device, small models run with TensorFlow Lite Micro, ESP-DL or STM32Cube.AI to classify signals, detect events or run computer vision on an ESP32-S3 or gateway, so only results need to leave the device. In the cloud, ML pipelines train on collected fleet data, and dashboards gain anomaly detection, forecasting or LLM features such as plain-language summaries of what a device or fleet has been doing.
Because the same team designs the hardware and writes the firmware, the integration is planned from both ends: the device sends what the platform and the models need, firmware updates can ship new models, and power, memory and bandwidth limits are respected rather than discovered late.
Technical Constraints and Trade-offs
- On-device models must fit the microcontroller's flash and RAM, so they are usually quantized to 8-bit and kept small; accuracy is traded against memory and inference time.
- Edge inference costs power. Battery devices typically run models on a duty cycle or on a trigger rather than continuously.
- Platform choice affects running cost and lock-in: managed platforms are fastest to start, while self-hosted stacks cost less at scale but need someone to operate them.
- Models trained on lab data often perform worse in the field, so a plan for collecting real device data and retraining is part of the design.
- LLM features need clear limits on what data is sent to a model provider, and their outputs should support decisions, not make safety-critical ones.
- Computer vision on devices depends on lighting, camera placement and frame rate, which are fixed by the hardware and enclosure, not only by the model.
Fabrication and Assembly Handoff
- Platform configuration and data model, exported or documented so it can be recreated
- Firmware repository with the integration code and the deployed model version
- Model files, training notebooks or pipeline code, and the dataset description
- Dashboard and alert definitions, plus API credentials handed over through your own accounts
- Data-flow diagram from device to platform to dashboard
From a real project
Three-phase smart energy meter, metering IC to cloud
A three-phase metering product built around an ATM90E36A metering IC and an ESP32-S3, with an EC200U cellular module and a CC2652P radio. Measurements travel from the device through an MQTT broker into time-series storage and dashboards, so the whole path from metering chip to cloud view was designed together.
Read the project on our PCB division's site (opens in a new tab)How an Electuni Engagement Works
Step 1
Discovery
Requirements analysis and feasibility study.
Step 2
Architecture
System design, component selection, and risk assessment.
Step 3
Design
Schematic, layout, firmware, and mechanical CAD.
Step 4
Review
DFM/DFT checks and design review with stakeholders.
Step 5
Prototype
Board fabrication, assembly, and firmware bring-up.
Step 6
Handoff
Production files, BOM, and full documentation package.
Tools & Technologies
- AWS IoT Core
- Azure IoT
- ThingsBoard
- Blynk
- Ubidots
- Grafana
- TimescaleDB
- InfluxDB
- Firebase
- Supabase
- Home Assistant
- TensorFlow Lite Micro
- ESP-DL
- STM32Cube.AI
- Python
AI & IoT Platform Integration FAQ
Should we use an IoT platform like ThingsBoard or build our own dashboard?
Use a platform such as ThingsBoard, Blynk or Ubidots when you need dashboards, device management and alerts quickly and their pricing fits your fleet. Build on Grafana with TimescaleDB, or a custom dashboard, when you need full control of data, branding or running cost at scale. Many products start on a platform and move later.
What is edge AI and when is it worth it?
Edge AI means running a machine learning model on the device itself instead of in the cloud. It is worth it when sending raw data is too costly in power or bandwidth, when the device must react without a connection, or when data should stay local. Typical examples are signal classification, event detection and simple computer vision.
Can an ESP32 or STM32 run a machine learning model?
Yes, for small models. ESP32-S3 and many STM32 parts run quantized neural networks with tools such as ESP-DL, TensorFlow Lite Micro and STM32Cube.AI. Typical workloads are keyword spotting, anomaly detection on sensor data, and simple image classification. Larger vision models usually need a gateway or a more capable processor.
How do you add LLM features to an IoT dashboard?
The dashboard sends a selected summary of device data, not raw streams, to a language model and shows the result, for example a plain-language summary of a device's week or an explanation of an alert. The design sets what data leaves your system, which provider is used, and how outputs are checked before anyone acts on them.
Can you connect devices to AWS IoT or Azure IoT that we already use?
Yes. Devices can be provisioned into your existing AWS IoT Core or Azure IoT account with per-device certificates, publish over MQTT to your topics, and feed your existing rules, storage and dashboards. Credentials and resources stay in your account, and the setup is documented so your team can run it.
Do we need lots of data before adding AI?
Not to start. A first model can often be trained on a modest dataset collected from prototypes, then improved as real field data arrives. Planning data collection from the first devices, with labels where possible, matters more than the initial dataset size, because field performance usually differs from lab results.