# SovAIHub Edge AI & TinyML assessment brief

Revision: 2026-09-26. Scope and acceptance worksheet, not a device certification.

## Business outcome

- One priority use case and operating owner:
- Baseline today (rules, inspection process, existing model):
- Cost of missed events, false alarms, and delayed decisions:
- Pilot site/device count and intended fleet size:

## Target device and environment

- MCU, gateway, industrial PC, accelerator, operating system / firmware:
- Flash, usable SRAM/RAM/VRAM, tensor arena, storage and buffers:
- Sensor/camera interfaces, sampling rate, acquisition window:
- Power budget, battery duty cycle, ambient temperature, cooling:
- Connectivity, offline duration, local identity and authorization:
- Data rights, retention, export restrictions, update channels:
- Access to physical devices and representative operating conditions:

## Data readiness

- Labeled examples, class imbalance, rare failures, and label reviewers:
- Sites, devices, lighting, operating modes, and environmental variation:
- Train/validation/holdout split by device, site, subject, item, or time:
- Controls preventing adjacent windows/frames leaking across splits:
- Sensitive-data handling and permission to train/adapt models:

## Acceptance contract

| Measure | Target | Test procedure / sample size | Evidence | Owner |
| --- | --- | --- | --- | --- |
| Fault/defect recall by class | | | | |
| False alarms per asset-day / false rejects | | | | |
| Acquisition-to-action p95 latency | | | | |
| Peak memory and storage | | | | |
| Energy per inference / battery life | | | | |
| Sustained thermal behavior | | | | |
| Sensor/network-loss behavior | | | | |
| Out-of-distribution / low-confidence fallback | | | | |
| Interrupted update and approved rollback | | | | |

## Milestones and go/no-go evidence

1. Feasibility: baseline, device constraints, data gaps, runtime/export compatibility.
2. Prototype: versioned model and preprocessing; quantization calibration; on-device parity.
3. Validation: holdout field results and sustained device measurements.
4. Rollout: artifact provenance, signed compatible update, canary, recovery, health checks.

## Ownership and commercial scope

- Deliverables, acceptance approvers, schedule, and fees:
- Hardware purchase, site work, travel, certifications, and support exclusions:
- Source/model artifact handover and license/redistribution review:
- Firmware, security updates, model drift, fleet recovery, and end-of-life owners:
- Independent safety controls and domain hazard analysis, if applicable:

Guide: https://sovaihub.com/solutions/edge-ai