By Shelly Tech · Engineering guide
A convincing demonstration is a useful starting point. The next step is to define what the complete device must do, under which conditions, and how success will be measured.
1. Define the physical task
Describe the decision the system must make and what happens next. Include the sensors, equipment, and people involved. This establishes the integration boundary around the model.
2. Choose where AI runs
For embedded AI and Edge AI, document the available device or gateway hardware, memory, power, and acceptable response time. For AI and IoT integration, also define what happens in cloud services and when connectivity is unavailable. Evaluate the model and application together under real operating conditions.
3. Define behavior when inputs fail
Decide what should happen when a sensor is unavailable, data is incomplete, or the model is uncertain. Make escalation and human oversight part of the product requirements.
4. Measure the complete path
Measure the path from sensor input through processing to the resulting action or interface. Model performance alone does not describe the response time of the whole product.
5. Agree on acceptance
Specify the scenarios, measurements, and evidence needed to accept the integrated system. A useful first conversation can start with these requirements even when the implementation is still open.