Embedded AI Validation Platform

Know what your embedded AI has actually proven.

EAIV turns firmware tests, TinyML benchmarks, sensor replays, power traces, and timing budgets into an auditable validation history. Run a mission, inspect the evidence, compare it with a known-good baseline, and see exactly what blocks the next release.

Built by Vers3Dynamics. Open source, hardware-aware, and runnable entirely in simulation.

Application areas: robotics, autonomous systems, industrial automation, medical devices, IoT, and edge-AI research.

EAIV Mission Control — release verdict, run status, budget headroom, and the prioritized diagnosis, over a guided workflow that goes from choosing a target to deciding whether to ship.

Capabilities

Subsystem	Function
Firmware testing	Flash, boot, and pattern-match firmware over serial, J-Link, QEMU, or an in-process simulated device
TinyML benchmarking	Latency distribution (p50/p90/p95/p99), throughput, startup time, output stability, model size, MAC estimate for .tflite/.onnx (mock fallback for dry runs)
Memory analysis	Static ELF ROM/RAM footprint, model flash cost, TFLite tensor-memory floor, budget gates
Power measurement	Mean/peak draw and energy per inference through the power_monitor plugin interface (simulated monitor plus INA226/PPK2 drivers; mock transports explicitly labelled)
Sensor fusion	Complementary, Mahony, Madgwick, 1-D Kalman, and 4-state EKF filters scored against ground-truth replay datasets (RMSE, drift)
Real-time profiling	WCET, jitter, and deadline-miss analysis from task telemetry
Hardware-in-the-loop	Deterministic fault injection (Gaussian noise, packet loss, timing jitter, sensor outages) over replayed or synthetic sensor streams
Telemetry	Typed serial protocol, per-board adapter plugins, capture to CSV, per-field statistics, sampling-rate and gap analysis
Evidence gates	Direction-aware comparisons that block failed suites, lost coverage, invalid numbers, target changes, and changed metric definitions; actionable CLI/dashboard explanations
Validation runs	Every run recorded with a manifest, event log, artifacts, and the resolved config that produced it; interrupted runs are detected, not left "running"
Diagnosis	Deterministic insight engine: evidence-backed findings with impact, confidence, and the next action — ordered by an explicit priority table
Measurement provenance	Every metric labelled measured, simulated, estimated, or mock, with its source (device, host, simulator, static analysis, dataset)
Release decision	Compatibility-checked run comparison (target, provenance, input hashes, suite coverage) with a recommendation and Markdown/JSON export
Mission Control	Guided Streamlit workflow: build a mission, launch it, watch it live, read the diagnosis, compare, promote a baseline

What a green gate means

A comparison passes only when the reports contain valid shared evidence, the candidate preserves baseline coverage, and no eligible metric worsens beyond the allowance. Each blocker names the affected suite or metric.

• Missing is a failure: removing a slow metric or skipping a previously executed suite cannot make a candidate look better.
• Definitions travel with measurements: declared units, directions, provenance, and sources must agree. Legacy reports remain readable.
• Promotion preserves trust: failed or malformed reports cannot replace a baseline; accepted snapshots are written atomically.
• Device readiness needs device evidence: host timings, mock transports, synthetic datasets, and unknown origins remain explicit. A passing simulation does not establish hardware readiness.

The verdict covers the executed workload and configured thresholds. It is not a certification or a guarantee outside that scope. See the evidence-gate contract for exact behavior and CI examples.

This is the beginning of the README. Read the rest at the original. Read the full README ↗