Laya Open-Source Decision Engine Outperforms Jev API

Independent developer NandhaKishorM has launched Laya, an Apache 2.0-licensed family of encoder-based decision models designed as a self-hosted alternative to TypeSafe's proprietary Jev API. Hosted on Hugging Face under the convaiinnovations organization, the project has already garnered over 4,000 GitHub stars. Unlike traditional autoregressive models, Laya operates as a non-autoregressive engine that processes typed questions against inputs like emails or JSON objects. It outputs structured probabilities in a single forward pass, eliminating the need for developers to parse messy prose or recover from malformed JSON.

In benchmark testing, Laya achieved a score of 0.766 on typed-decisions, surpassing Jev's score of 0.727. Crucially, Laya operates roughly 7.8 times faster than Jev at p50 latency, clocking a single-question forward pass at 33 milliseconds on an Nvidia T4 GPU, which drops to 7.2 milliseconds per question when batched. However, the model struggles with high-cardinality choices, scoring just 0.425 on the Banking77 benchmark compared to Jev's 0.870, a limitation caused by the token budget allocated per label.

The system includes three distinct checkpoints and an automated router that detects scripts and languages in under 0.5 milliseconds to dispatch requests. The English checkpoint utilizes ModernBERT-large with 421 million parameters and a 512-token context window. For global applications, the multilingual checkpoint uses mmBERT-base with 322 million parameters, supporting over 100 languages with a 1,024-token context window and offering twice the throughput of the English version. A third checkpoint is specifically tuned for typed-decision tasks.

For machine learning practitioners, Laya offers a highly efficient, pip-installable solution for routing tickets, detecting intent, and selecting models. Because it runs on self-hosted encoders, teams can bypass external API costs and latency. Training is highly accessible, with a fine-tuning notebook capable of running on Kaggle's free tier using two Nvidia T4 GPUs. While production deployments still require schema validation and timeout handling, Laya provides a fast, predictable foundation for branching code paths.