ABOUT
NeuroFS is a stateful semantic routing engine for LLM applications. We believe AI routing should be learned, not hardcoded — adapting to each user over time, not guessing from static rules.
The Problem
Every LLM call burns tokens on context the model doesn't need. Static routing sends every prompt through the same path regardless of what the user actually wants.
Token costs scale linearly. Latency compounds. Most "RAG" pipelines retrieve everything and hope the model figures it out.
The Solution
NeuroFS decomposes every prompt across 7 semantic dimensions and routes it through a learned 100³ activation cube. Graph-spread and Hebbian learning adapt routing per user over time.
The result: precise adapter selection, surgical tool routing, and KV-cache prefetch — all in under 2ms, with 96.7% routing accuracy.
Core Architecture
Semantic Cube
100³ sparse activation space. RoaringBitmap storage. Three axes: Domain, Intent, Style.
Graph Engine
BFS spread activation over a typed semantic graph. Edges weighted by co-activation history.
Decay Engine
Exponential decay with 1-hour half-life. Regions fade unless reinforced by new interactions.
Learning Engine
Per-user Hebbian learning refines edge weights. Population aggregation generalizes across users.
7 Routing Dimensions
Ready to try it?
Sign in, generate an API key, and route your first prompt in under a minute.