About the hackathon
Inference Starts Here was a vLLM and llm-d hackathon run by The Open Accelerator, Red Hat's Boston AI hub in the Seaport, and mentored by vLLM core maintainers. The theme was serving open models efficiently rather than calling a hosted API, across six challenge tracks and three skill lanes.
The track we competed in
We entered the Agentic Edge track, powered by NemoClaw, in the Deep Tech lane. It asked for agent systems that run on local inference with real guardrails, not a thin wrapper around a cloud endpoint, which pushed us to keep the whole pipeline on vLLM.
What we built
PRGenie triages GitHub pull requests with seven cooperating agents that score author trust, flag risky paths, extract intent, and route the review. It runs on vLLM with NVIDIA Nemotron, and every action passes through a policy runtime so the model can only act within defined limits.
What we won
We won the Agentic Edge track in the Deep Tech lane, and took home an NVIDIA GeForce RTX 4090 for the build.