Haiyang Wu

Co-Founder & CTO, Krewva

I build AI systems people can actually trust.

Currently Co-Founder & CTO of Krewva — the AI for your relationships. It reads across the apps you already use, surfaces the obligations hiding in your conversations, drafts the action in your voice, and waits for your yes.

Previously founded Machina AI (zero-cloud visual inspection for high-mix manufacturing) and open-source SMBoost, a reliability harness for small local LLMs.

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 reliably in the real world, under load, with no second chances.

Today I'm Co-Founder & CTO of Krewva — the AI for your relationships. Your mom's on WhatsApp, your kid's school sends iMessages, your boss pings on Slack, and no app sees across the others. Krewva is the one place they meet: it reads the context, surfaces the obligation buried three weeks deep, drafts the reply in your voice, and acts only on your approval. I own the technical bet — a system that interprets, decides, and acts across fragmented apps, reliably enough that people trust it with the relationships that matter most.

I got here through hard-won lessons. In 2025 I founded Machina AI to bring computer-vision inspection to high-mix, low-volume factories that can't send production-line imagery to the cloud. I built EdgeRunner 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 turned out to be a harder problem than the engineering.

The deeper lesson sat below the stack: 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 became SMBoost, and that same obsession — making models trustworthy, not just capable — is exactly what Krewva runs on.

Currently

Projects

Krewva— krewva.com

The AI for your relationships. Reads across iMessage, WhatsApp, Gmail, Calendar, and Slack; surfaces what you've promised; drafts actions in your voice; acts only with your approval.

Role: Co-Founder & CTO — leading engineering and core architecture (interpretation → decision → action across fragmented apps).

EdgeRunner— haiyang5535/edge-runner

Real-time edge AI vision system on NVIDIA Jetson Orin Nano combining YOLO detection with VLM reasoning.

Stack: TensorRT FP16, ByteTrack, GBNF-constrained VLM JSON output, dual-thread architecture.

Results: <12ms P99 detection latency on Jetson Orin Nano. Signed pilot LOI with a Shenzhen powder metallurgy manufacturer.

SMBoost— haiyang5535/smboost

Open-source reliability harness for 2B–7B local LLMs in production.

Origin: came directly out of a failure mode in the Machina pilot — small VLMs on Jetson worked in the median but broke in the tail, and that tail was the adoption blocker, not raw capability.

Stack: hierarchical state machine, adaptive robustness scorer, GBNF grammar constraints, Pydantic invariants.

Stack

Python · NVIDIA Jetson · TensorRT · YOLO · VLMs · Computer Vision · ML Systems · Edge Deployment

Contact