Edge AI: artificial intelligence inside the product
ProtoMakers develops electronic products with Edge AI for companies in Israel: models that run on the microcontroller or on a small computer inside the product, with no dependency on the cloud. We build the whole chain, from the sensors and data collection, through training the model and compressing it to fit the hardware, to the firmware that runs it.
What we do with Edge AI
- A feasibility check: does AI really add value to the product, and which model fits the hardware
- Collecting and labeling sensor data
- Fault and anomaly detection and predictive maintenance from vibration, current, voltage and temperature
- Sound and signal classification, and motion recognition from IMU sensors
- Image processing on a small Linux computer or on the ESP32-S3
- Model compression (quantization) so models fit in memory and meet run-time limits
- Running models on STM32, ESP32-S3 or a small Linux computer, with tools such as TensorFlow Lite for Microcontrollers, STM32Cube.AI and Edge Impulse
- Measuring accuracy, run time, memory and power consumption on the real hardware
When Edge AI is the right choice
- You need an immediate decision, without waiting for an answer from the cloud
- There is no continuous internet connection, for example in the field or inside a machine
- The data is sensitive and is better kept on the device
- Sending all the raw data to the cloud costs too much in communication or battery
When something else fits better
- A simple rule or a fixed threshold solves the problem: no AI is needed, and that is both cheaper and more reliable. If so, we will tell you during the feasibility check
- The model is very large, such as a large language model (LLM): it is better to run it in the cloud and connect the product to it
- There is no real field data yet: first collect data from a prototype, and only then train
- You need a research team developing new kinds of models from scratch: a large company or a dedicated data science team is a better fit
Who does the work
The same engineer is responsible for the sensor, the board, the firmware and integrating the model. With Edge AI this matters a lot, because a model that worked perfectly on a PC often fails on the device: a different sampling rate, noise from the power supply, too little memory or too long a run time. With one engineer responsible for the whole chain, there is no argument between the data team and the embedded team about whose problem it is. And in a small company you talk directly to the engineer, with no project manager in between, which is both faster and cheaper.
FAQ
Who develops Edge AI for electronic products in Israel?
ProtoMakers, a development company in Kiryat Tivon, develops electronic products with AI running inside the device, for companies across Israel. One engineer is responsible for the sensors, the board, the firmware and integrating the model.
What is the difference between Edge AI and cloud AI?
With Edge AI the model runs inside the device itself, so the answer is immediate, the product works without internet, and the data stays on the device. Cloud AI can run much larger models, but it needs a continuous connection and adds delay and cost to every use.
Can AI run on a microcontroller?
Yes, small models. Anomaly detection, sound classification and motion recognition run well on microcontrollers such as STM32 and ESP32-S3. Image processing usually needs a stronger processor, such as a small Linux computer.
How much data do we need to start?
It depends on the task. Usually we start by collecting data from a prototype under real conditions. For anomaly detection, data from normal operation alone is sometimes enough, because the model learns what is normal and flags anything that deviates.
Who owns the model and the code at the end?
You do. The source code, the model, the training scripts and the collected data are handed over at the end of the project.
Have a project on the table?
A free 30-minute intro call, no obligation.