Jev-style decision models multiply as researcher says idea predates TypeSafe release

A claim arrives after the launch

Decision model APIs modeled on TypeSafe's Jev moved quickly from single product to crowded category. The release was followed by Microsoft Decision, OpenAI Decisions API and Fastino Glide, while Cloudflare Clef and Perplexity Decider arrived with open weights for users who can run them on their own hardware.

The dispute centers on Nandakishor Mukkunnoth, founder of ConvAI Innovations. He said he designed the Jev architecture one year back and pointed to an arXiv paper on SalesRLAgent that he published on March 30, 2025, along with model weights and dataset. He later published a September paper that more closely aligns with how Jev operates.

What the models actually do

Jev and similar tools are APIs driven by decision models for use in your own code. A developer can feed input such as a customer's online shopping transactions together with questions, including a yes or no return question and a multiple choice item question. The model returns a probability for each answer, which can then be used in if statements.

The practical application is described as novel even if the concept is not strictly new. The appeal is that the system avoids token by token text generation, conversations, repeated conversation history, large context loads and conversational hallucination. Instead it answers questions in parallel with probabilities, and the source describes it as very fast and much cheaper than standard LLMs.

Open projects keep pace

TypeSafe paired Jev with a brief description of its System One architecture and a complete hosted product with servers, API, documentation, developer kits and pricing. Mukkunnoth said his earlier work was overlooked because it was built for a vertical use case, while Jev was made a general purpose decision engine. He said he chose not to remain bitter and released the open source Laya around the same day.

Laya is presented as a Jev style model and interface that can run on local hardware, with multi language support. Unlike many followers that start from Alibaba's open source Qwen and add layers for input, output and decision making, Laya uses encoder only models from the Bert family. The source notes that the speed of the alternatives arriving makes it easy to imagine that some may have borrowed at least a few notes from his work.

Sources: Tomshardware