Executive summary

EdgeSutra AI 0.2.1 is an evidence-first toolkit that tells engineering teams what is proven, what is merely declared and what is unknown when they plan to run an AI model on an edge device (phone, Raspberry Pi, embedded board, laptop NPU).

The problem

Teams routinely ship quantized models on the strength of a file size, a vendor datasheet or a laptop benchmark. Those are not evidence that a model runs correctly, fast, within memory and within power budget on the real target. The cost of a wrong assumption is a failed field rollout.

What EdgeSutra does

CapabilityOutcome for a decision maker
One-command audit (edgesutra audit)A shareable HTML/Markdown/JSON report with a verdict and a list of what was not tested
Static compatibility checkEarly, cheap rejection of impossible model/runtime/device combinations
Safe ONNX metadata inspectionOperators, opsets and SHA-256 read without ever executing the model
Host-only benchmark with guardrailsReal timings, labelled as host-only; never passed off as device latency
20 portable agent skillsThe same review discipline inside Claude Code, Cursor, Codex, Copilot, Gemini CLI

What it deliberately does not do

It is not an inference engine and does not run models on devices yet. It never invents energy, carbon, memory-peak or device-latency figures. It cannot output "ready to ship": the best verdict is REVIEW_REQUIRED until on-device evidence and a human sign-off exist.

Why trust it

Status

Working offline core (v0.2.1). Planned: ONNX Runtime CPU inference, isolated quantization, first real-device connector. See Roadmap.

Creator: Vatsal Shah · Repository: https://github.com/shahvatsalofficial/edgesutra-ai