6954

Evidence Engine

Things I built, measured,
and published.

Independent work in Edge AI · Distributed Systems · AI Products · Productization.

This is not a portfolio of screenshots. Each entry states the question it set out to answer, the numbers it actually measured, and what those numbers do not prove.

Explore workRead the notes

Featured work

Open source · Edge NPU Cluster

Can three 6-TOPS NPUs scale inference throughput 3×?

421 measurements · 3.00× scale-out · RK3576

Three NanoPi R76S boards, one scheduler, plain gRPC over ordinary Ethernet. 421 valid hardware runs with zero inference errors. It scales — and four separate findings almost stopped it.

Explore NPUDureEvidenceGitHub

Product · live

Run AI on your own device and compete on score

AI Arcade runs models on your own hardware and turns the result into a score you can compare.

Open AI Arcade

Latest note

Note · 2026-09-02

Edge NPU Cluster: Do Three 6-TOPS NPUs Really Make 18 TOPS?

421 hardware measurements across three RK3576 nodes over standard 2.5GbE and gRPC.

Read한국어

한국어

The same work in Korean. 주요 글은 한국어로도 읽을 수 있습니다.

NPUDure소개

In preparation

Each project is published as its own event rather than all at once. Crescendo and Pulse (research), Canto (product) and Vending AI (case study) follow NPUDure.