"FB team on AMD Developer Hackathon: ACT II Hackathon"

Team Idea

AI agent that benchmarks PyTorch/LLM workloads on AMD GPUs and recommends faster ROCm deployment settings.

Team Members

Som Tripathi

somtri

Submission

flashbulb-amd: zero-token agent memory on AMD

Self-hostable memory for AI agents. A neural surprise gate decides what to keep for zero LLM tokens, and nothing leaves your machine. Memory and a Gemma 4 model run on one AMD MI300X, ROCm end to end. Free, open source for devs, sovereign for enterprise.