'Missing context' big challenge for scaling AI in India: Uniphore cofounder Ravi Saraogi

For Indian use cases, large language models (LLMs) need to understand context rather than rely only on voice-to-text conversion, said Ravi Saraogi, cofounder of enterprise AI company Uniphore. He highlighted the ‘missing context’ as a key challenge in building scalable AI systems.

Uniphore achieved the unicorn tag in 2022 and closed a $260 million funding round in 2025 from US chip majors Nvidia, AMD, cloud data storage company Snowflake, and data analytics firm Databricks.

“When we talk about voice-first, it’s important to understand the difference between simply converting voice to text and actually extracting meaning from what is being said,” Saraogi said while speaking at ET’s AI Conclave & Awards 2025 on Thursday.


He added that many times the Indians communicate through “mix of languages, or with implied meaning.”

Speaking about the need for context-aware systems Avataar.ai’s Sravanth Aluru, cofounder and CEO, said for India, that challenge remains unique.

“Building AI for the next 500 million users requires low-latency, cost-efficient models that work across dialects and languages, and often through voice rather than text.” he said adding that one can’t assume prompt typing like we do in Bengaluru (metro cities).

“You need multi-dialect, voice-first intelligence, with humans in the loop to reinforce learning,”Aluru added. Avataar.ai works in the domain of developing interactive 3D and AR experiences for e-commerce brands.

The two spoke about the need to distinguish between data dump and data moat, highlighting that Indian enterprises are sitting on vast amounts of underutilised data that could become a long-term ‘AI moat’ if structured well.

Saraogi highlighted that behavioural data is an ignored asset. “How a rural banking customer pauses before agreeing to a loan, whether silence indicates doubt or consideration, even how Indians nod their heads differently across states-these are data points,” he said.

According to him, enterprises that can convert such contextual signals into proprietary intelligence can build defensible IP and monetise it across markets. “That’s how Indian companies can create a real data moat-not just by scale, but by extracting knowledge from it,” he said.