Agent infrastructure · State of the art · Open source
Laya-Decision
A small, fast engine that answers structured questions — yes/no, pick-one, give-a-score — in a single step, without writing any text. Written in Rust, runs on your own machine.
§ What it is
Laya-Decision is a pure-Rust port of Laya, a multilingual non-autoregressive decision engine. You hand it a state (text, JSON, or a conversation) and typed questions — choice, score, noul — and it returns scored answers in one pass. It’s the clearest expression of a conviction we keep proving: the model is not the product. Where a system would otherwise spend an LLM call, Laya makes a fast, local, deterministic decision instead.
§ Why it matters
- Fast and local. Inference runs natively on candle — CPU or Apple-Silicon Metal. Routing, language detection, and email cleaning have a weight-free, zero-dependency mode.
- Faithful. The candle backend matches the upstream PyTorch model to within 1e-4 across all three checkpoints.
- Reusable. Published to crates.io under Apache-2.0, with an HTTP server and CLI — the decision layer that gates and routes inside Immediacy, now a building block anyone can use.
§ Use it
laya-decision = "0.1"