Team Idea
We are building EdgePower, a hybrid AI inference runtime that routes sensor events between a local LLM (phi-3-mini via Ollama) and cloud Gemini based on per-event confidence and severity — suppressing noise locally, escalating only when necessary, achieving >95% bandwidth reduction. This targets the core physical AI infrastructure problem: reliable, cost-aware inference under intermittent connectivity, architected for SiMa.ai Modalix MLSoC-class deployments. Stack: Redis Streams, Ollama phi-3-mini, a Routing Policy Engine, SQLite WAL with causal replay, Qdrant RAG, Gemini via Lobster Trap governance proxy, and OpenTelemetry per-inference tracing. Demo: live two-site operation with OTel cost/latency observability, a CRITICAL event triggering RAG→Gemini→dashboard diagnosis, and a network partition where Site A classifies locally, accumulates replay state, reconnects, and flushes causally — never losing an event, never stopping.
Team Members
Jay Liu
JayLiu
Code Design PMF
.png&w=256&q=75)
