Positioning A multilingual, non-autoregressive decision engine that outputs typed decisions (classification, scoring, yes/no judgment) for a text state in a single forward pass, replacing generative LLMs with fast, deterministic inference — an answer comes back as a typed decision rather than generated text.
What it does Decision needs are specified as typed questions over the input state; a single forward pass returns the answer with a calibrated confidence, with no text generation and no hallucination. A router detects language and script in sub-millisecond time and automatically routes among three checkpoints, covering 100+ languages across those checkpoints. The project reports (self-reported) that fine-tuning on your own decision data lifts accuracy on its benchmark from 0.362 to 0.766. It also ships a Jev-compatible HTTP server, MCP and LangChain integrations, a CLI, and a web playground, so the same engine can be reached from an agent stack, a shell, or a browser.
Characteristics Built on ModernBERT and mmBERT encoders, trained with reinforcement learning under a strictly proper scoring rule; project-reported latency is about 33 ms per question on a T4, or 7.2 ms per question in batch. Long documents up to 8192 tokens are supported, with optional ONNX and TileLang GPU acceleration paths and Docker and NixOS deployment recipes.
When to use Suited to high-throughput structured decisions such as ticket triage, email routing, content moderation, and churn early warning, especially multilingual production environments — where the router picks the checkpoint per input — and business that cannot tolerate hallucination. Not suited to open-ended text generation, multi-step complex reasoning, or deep-comprehension needs beyond its long-document limits.