ConvAI Founder Alleges TypeSafe Copied Jev Architecture Before Market Flood
A researcher claims TypeSafe’s Jev decision model architecture was derived from his prior work, sparking a dispute as similar tools proliferate.
Key points
- Nandakishor Mukkunnoth of ConvAI Innovations states he designed the Jev architecture a year before TypeSafe's release.
- TypeSafe launched Jev as a hosted product with APIs, documentation, and pricing, differing from the original vertical use case.
- The architecture uses decision models to output probabilities for specific questions, avoiding standard LLM token generation.
Nandakishor Mukkunnoth, founder of ConvAI Innovations, alleges that TypeSafe copied the architecture for its Jev decision model after he published the underlying concepts more than a year earlier, according to a report by Tom's Hardware. This claim has emerged as numerous AI companies and open-source projects have released similar decision-model APIs following TypeSafe’s launch.
Mukkunnoth detailed his position in a blog post released days after TypeSafe introduced Jev and its System One architecture. He stated that he designed the core notions for the Jev architecture one year prior to TypeSafe’s release. He referenced an arXiv paper on SalesRLAgent, published on March 30, 2025, which described a model for probability estimation. Alongside that paper, he released model weights and a dataset.
Context of the Dispute
The core of the disagreement centers on the scope and presentation of the technology. Mukkunnoth explained that his original work, SalesRLAgent, was built for a specific vertical use case. He expressed frustration that his contribution was overlooked because it addressed a narrow problem rather than a broad one. TypeSafe, by contrast, released Jev as a general-purpose decision engine.
TypeSafe packaged Jev as a complete, hosted product. The offering includes its own servers, a public API, developer documentation, and developer kits. The company also established pricing for access to the service. Mukkunnoth noted that he later published a September paper that more closely aligns with the operational mechanics of Jev, further highlighting the similarities between his earlier research and the commercial product.
How the Technology Works
Jev and its emerging clones operate as APIs driven by decision models intended for integration into custom code. Users feed input data, such as customer online shopping transactions, alongside specific questions to the model. For example, a user might ask if a customer is looking for a return or which item they are referring to.
The model answers these questions or multiple-choice options with a probability score. Developers can then use these probabilities in their own conditional logic statements. This approach sidesteps common issues associated with large language models, such as token-by-token text generation and conversational hallucinations. It eliminates the need for re-feeding conversation history or providing large amounts of contextual information.
This method is noted for its speed and cost efficiency compared to standard LLMs. The models answer questions in parallel, providing direct data in the form of probabilities rather than generating natural language responses. This practical application is considered novel, even if the underlying concept of probability estimation is not strictly new.
What should you do?
Security and technology professionals should monitor the development of these decision-model APIs for potential vulnerabilities in how they process input data. Since these models output probabilities for code integration, teams should verify the reliability and accuracy of the outputs in their specific contexts. Watch for updates regarding the dispute between ConvAI Innovations and TypeSafe, as it may impact the stability or licensing of these tools.
Background: Large language models
Large language models are a specific type of neural network trained on text. Generative AI is the broader umbrella term for any system that creates new content, including images and code. Not all generative AI uses large language models, but most modern text generators do.
What to do and how to stay safe: TypeSafe
- Review any existing integrations with Jev-like decision models to ensure they handle probability outputs securely within your application logic.
- Monitor official channels from TypeSafe and ConvAI Innovations for statements regarding the architectural dispute and any resulting technical changes.
- Evaluate alternative decision-model providers to reduce dependency on a single vendor while the origin dispute remains unresolved.
- Ensure that input data fed into these models does not contain sensitive information that could be exposed through probability outputs or logging.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is the specific claim made by ConvAI Innovations against TypeSafe?
Nandakishor Mukkunnoth claims he designed the Jev architecture one year before TypeSafe released it, citing his March 2025 arXiv paper on SalesRLAgent.
How does the Jev architecture differ from standard large language models?
Jev uses decision models to output probabilities for specific questions, avoiding token-by-token text generation, conversation history, and hallucinations.
When was the original research paper cited by the accuser published?
The arXiv paper on SalesRLAgent, which described the probability estimation model, was published on March 30, 2025.
