THIRUVANANTHAPURAM : Their Watasale a machine learning-powered physical retail store had been a hit globally, bearing testament to Subhash Sasidharakurup, Dileep Jacob and Vinci Mathews’ expertise in AI. The team has now come up with India’s first AI-powered search engine that is designed to support local languages in India and other countries.
Detailing what Nofrills AI entails, Vinci Mathews, cofounder of the startup, says, “This platform represents a milestone in making information discovery much more easier by just having a conversation. This makes it more accessible and user-friendly for regional language users.” The team’s first venture Watasale had been acquired by Amazon.
He says Nofrills AI is an AI-powered discovery engine that provides real-time, accurate answers directly from the web, supporting both English and local languages seamlessly.
“Users can have conversations in their language preferences, ensuring that responses are tailored to their preference. This is made possible through the platform’s hybrid technology and WISE architecture, which ensure clarity and precision in local languages,” says Vinci.
Another highlight of Nofrills AI is that it is the first live LLM model from India to support local languages fully. “Another plus is that the platform is available for free to all users. It has been designed to make information discovery easy and intuitive, facilitating meaningful actions based on the discovered information. These innovations enable accurate and efficient responses, even in complex local languages like Malayalam,” he adds.
“Nofrills AI focuses on essential intelligence features to understand and interpret live web data from credible and authentic sources. Unlike global players who restrict access to strong models due to high running costs, WISE reduces parametric memory load by segregating intelligence from data. This architecture, built using open weights and fine-tuned for improved reading comprehension, can interpret live data and evaluate genuine, credible, and even sarcastic sources better than its peers while maintaining small size and lower operational costs.”
Vinci says, “One of the primary obstacles that LLM-based applications face when implementing local language support is the significant decline in accuracy resulting from the larger token size of local languages. For example, Malayalam, a language that requires roughly ten times the number of tokens for a given sentence compared to English, poses a significant challenge. This increased load can have adverse effects on both training and inference, ultimately impacting model accuracy in local languages.”
To address this challenge, the startup developed a hybrid approach using a neuro-symbolic framework. “Our architecture estimates the output size, and if it falls below a safe threshold, the model directly generates responses in the local language. Otherwise, it invokes an intermediary translation layer. This innovative strategy ensures that accuracy in local languages is comparable to that of English, significantly enhancing user experience,” says Vinci.
The team has big ambitions. “Nofrills AI operates out of both India and the US, with plans to expand globally,” says Vinci. Backed by angel investors, the startup’s product truly stands up to its name Nofrills. “The name means without unnecessary extras, especially ones for decoration,” says the co-founder. The rest of the team is comprised of experienced professionals from the Watasale and Amazon who are dedicated to improving information discovery across various sectors.