Krutrim AI models improving almost every week, says company executive

The startup aims to address the inherent hallucination issue in AI models by leveraging user feedback from its beta program

Krutrim, the AI start-up launched by Ola founder Bhavish Aggarwal, faced criticism from users over translation inaccuracies, factual errors, and lapses in logical reasoning. Krutrim rolled out its public beta in February. Senior executives, however, said that the platform is seeing improvement almost every week.

“Any AI model will have a hallucination problem. It’s a probabilistic model. This is the entire reason why we are doing the beta programme, to get user feedback and use that data to improve the model over time. Almost every week, we are seeing an improvement in all the issues that consumers are pointing out,” said Ravi Jain, Head - Strategy, Krutrim on the sidelines of the Business Standard Manthan event in New Delhi.

Krutrim is investing heavily in the proper training of its models, using factual and relevant data sets to mitigate inaccuracies, said Jain.

“We are looking at all relevant and factual Indian open web data to train our models. One of our key pillars is the quality, the quantum and the contextuality of the data, and we are making investments in those areas,” he said.

Krutrim became India’s first unicorn — a term used to describe start-ups valued at $1 billion or above — this year in January, after it had raised its inaugural funding round worth $50 million from investors such as Matrix Partners India.

The company also claimed the distinction of being India’s first AI company to reach unicorn status and the title of the country’s fastest unicorn. Krutrim, which translates to ‘artificial’ in Sanskrit, was launched in April 2023 by Aggarwal and Krishnamurthy Venugopala Tenneti, a board member of ANI Technologies which owns Ola and Ola Electric.

When asked about some of the notable use cases derived from feedback received during the public beta, Jain said the usage exhibits interesting trends. “We are seeing very interesting use cases across areas like education and content creation. In Tier-2 and Tier-3 regions, we are seeing a lot of multilingual applications. For example, government employees using it to write letters in English, or even small business owners creating their catalogues,” Jain said.

Krutrim currently supports over ten Indian languages including English, Hindi, Tamil, Bengali, Marathi, Kannada, Gujarati and ‘Hinglish’.

The company is now working on developing more sophisticated models for its platform and create compute infrastructure. “We need to move fast... We need more people and resources, more compute to deploy to come up to a level where we can be satisfied,” Jain said.

This comes at a time when competition in the sector is heating up. Homegrown generative AI start-up Sarvam AI, which not too long ago raised $41 million in Series A funding, recently unveiled OpenHathi, the first Hindi large language model.

Meanwhile, Reliance Industries, in collaboration with nine Indian Institutes of Technology (IITs), also introduced its own AI model BharatGPT Hanooman to address key challenges in AI development and aiming a leadership position in the sector.

“The space is too big. And we need a lot more ideas and efforts. And that is what will take this space, and our country in particular, forward... I am sure many such efforts will start in due course. As a company, we have to be at the cutting edge. And that is our responsibility, and our focus right now,” Jain said.