Haiyang Wu

About

I came to engineering through mathematics and statistics at UConn, graduating in December 2025. The path into ML and systems work was self-directed — training models that performed well in notebooks, then learning the harder half: running them on constrained hardware, under load, with no second chances.

In 2025 I founded Machina AI to bring computer-vision quality inspection to high-mix, low-volume factories that can't — or won't — send production-line imagery to the cloud. I built EdgeRunner, a real-time vision pipeline combining YOLO detection, ByteTrack, and a small VLM on a Jetson Orin Nano, and took it to a Shenzhen powder metallurgy manufacturer. We signed a pilot LOI. We did not sign a contract. ToB sales on a single-founder timeline is a different problem than I had framed it going in; the engineering was the easy half.

The harder problem sat below the stack I was trying to sell. Running a 3B VLM on the Jetson, the model worked when it worked — but its tail behavior (malformed JSON, hallucinated classes, unjustified confidence) was what kept a plant from trusting it. Not raw capability. That observation became SMBoost: a reliability harness for small LLMs in production — grammar-constrained outputs, a hierarchical state machine around inference, invariant checks on structured results, and a scoring layer that flags silent degradation.

I'm looking for the next room to build in — either a team pushing Edge AI into production at real scale, or a continuation of SMBoost as a standalone project.