This Kerala Engineer Built Open-Source Jev Alternative a Year Before the Hype

Lately, the internet has been abuzz with a new, non-autoregressive AI model developed by TypeSafe AI. Jev doesn’t generate text token-by-token. Instead, it processes the entire payload in a single, parallel evaluation pass.

But Kerala-based engineer and founder of ConvAI Innovations, Nandakishor M, had already been exploring the idea months before Jev’s launch. His project, Laya, is a small decision-making model built around the idea that not every AI task needs a large language model. He began working on Laya in April last year while exploring ways to identify hallucinations before an LLM finishes generating its response. 

Nandakishor does not want Laya to become a closed model. Unlike Jev, which charges $0.042 per million input tokens, Laya can be self-hosted without any fee.

The model is open-source under an Apache 2.0 licence, allowing developers to build on top of it. The model has attracted more than 3,500 likes on Hugging Face and has continued to trend on the platform. It was even the top trending model on Hugging Face, with CEO Clem Delangue praising the model.

“I just open-sourced it so everyone can work on it and build stuff on it,” Nandakishor tells AIM.

He also sees Laya as an early open-source implementation of the calibrated decision approach he has been researching. His ambition is not to position the model as a direct replacement for Jev or large language models, but as an open foundation that developers can fine-tune for specific decision-making workloads.

Decisions Over Generating Text

Nandakishor argues that transformer-based LLMs have a fundamental limitation when it comes to hallucinations.

“These models tend to hallucinate,” he says, pointing to situations where a model may generate an answer even when without any reliable information in its training data.

His initial research examined whether hallucinations could be detected in models’ internal representations before they generated the final text. The work led to a paper called Confidence Aware Routing, based on a contrastive learning approach.

Laya takes the idea further as a Type 1 decision model. Instead of generating tokens one after another, it produces a decision, score, or probability.

The architecture combines a bidirectional transformer representation with additional layers for classification. Its training pipeline uses Reinforcement Learning for Calibrated Decision, or RLCD. The post-training AI method rewards models for producing statistically accurate confidence probabilities alongside typed decisions.

Laya’s multilingual base model supports a maximum sequence length of 1,024 tokens, which can be extended to 8,192 tokens. Nandakishor says this is sufficient for most use cases.

Laya supports up to 100 languages, including Hindi, Mandarin, and Japanese.

Nandakishor says the objective is to improve the model’s confidence to match the actual probability of an outcome. Say the model predicts with 90% confidence that an email is spam, but the email is actually genuine. The model is penalised for the mistake. 

From Phishing Detection to Model Routing

Laya can be used for tasks where a binary or scored decision is sufficient.

Nandakishor sees applications for Laya in phishing detection, prompt-injection detection, content moderation, model routing, and customer-support workflows.

In a real-world scenario, a customer asking for a refund may not require a large language model to process the request. A smaller decision model could identify the intent and route the request directly.

Laya can also act as a router between larger models, deciding which model or workflow should handle a particular request.

Nandakishor says the model has been benchmarked up to around 7,000 tokens, although he sees around 6,000 tokens as a practical range. He claims inference can take around 30–35 milliseconds, making it 6–8x faster than Jev.

In the LocalLlaMA typed-decisions benchmark evaluating AI models on output structure data over 2,000 decisions, Laya scored 0.766, compared with 0.727 for Jev 1.13.0, giving Laya a 3.9% higher score and putting it above the 0.735 teacher benchmark, which indicates that the model memorises noise, edge cases, and flukes in the training data.

Edge Use

Laya has around 1 million parameters and can run on devices ranging from Raspberry Pi and smartphones to laptops and embedded systems.

Nandakishor calls this a ‘zero-egress’ approach because sensitive information can remain on the device rather than being sent to a cloud service.

He also sees edge AI as a way to reduce latency, cloud dependency, and infrastructure costs.

Decision models can also reduce LLM spend. For a company spending around $5,000 a month on LLM APIs, he estimates that routing some workloads through a smaller decision model could potentially bring that down to between $1,000 and $2,000, depending on the use case.

He stresses that the actual savings would depend on how many decisions can be handled without invoking a larger model.

Agentic workflows are on the rise and they increasingly require fast and inexpensive decisions. That’s where Laya can work best.

Laya Max Coming Soon

Nandakishor says the next version, Laya Max, is under development.

ConvAI Innovations has been working with researchers and collaborators and is exploring partnerships around the project. The founder does not currently plan to raise funding specifically for Laya. He wants to keep the research open while allowing commercial partners to build products around it.

From Engineering Degree to AI Research

Nandakishor’s path to Laya did not begin with a computer science degree.

He studied electrical engineering at Government Engineering College, Kannur. Nandakishor eventually secured admission to IISER Mohali but could not afford the course.

“Back in 2008, our family made only ₹200–300 per day, and we were barely surviving,” he recalls.

During engineering, he became involved in hackathons, conferences, and technology initiatives. In his second year, he filed a provisional patent for a theoretical graphene-based hybrid solar panel, partly to understand how patents, NDAs and intellectual property worked.

His next focus is edge computing. Large data centres have a massive environmental and electrical footprint. Rather than building another large model that depends on massive infrastructure, Nandakishor wants to explore smaller systems that can make decisions locally.