Edge AI & TinyML Readiness Checklist

Device type, CPU/GPU/NPU, memory, storage, OS, and runtime support are known.

You know whether the task needs real-time, near-real-time, or batch inference.

Battery, heat, fanless operation, and industrial enclosure constraints are considered.

Sensor, image, audio, event, or machine data exists for the exact operating condition.

There is a way to train or validate the model against real success/failure examples.

Seasonality, machine aging, lighting, sensor variation, and environment changes are considered.

Classification, anomaly detection, visual inspection, wake-word, sensor prediction, or compact inference is clearly defined.

TFLite, ONNX, GGUF, TensorRT, OpenVINO, vendor SDK, or another device runtime is identified.

Accuracy, latency, memory, CPU/GPU usage, and failure cases will be tested on the real device.

The solution knows when it can sync, when it must work offline, and how results are buffered.

Model versions, signatures, checksums, approvals, rollback, and deployment records are managed.

Device health, inference failures, drift indicators, and model version are observable.

Access, tampering, local storage, logs, and model extraction risks are reviewed.

Uncertain predictions, alarms, or safety-impacting decisions have a clear escalation flow.