The world is in a headlong rush towards AI, and the spotlight remains fixed on tech giants in the US and China. Conversations there often circle around scaling models, expanding compute power and securing vast amounts of data. 

Yet, in India, a quieter current is beginning to flow. Here, researchers and startups are approaching the challenge differently—focusing less on scale and more on efficiency, inclusivity and domain-specific needs. 

For Shunya Labs, the central question isn’t “how big can the model be trained?” but “how well can it reason?” Sourav Banerjee, co-founder and technical architect, is blunt about the shortcomings of today’s large language models. “They mimic the act of reasoning, but they don’t actually reason,” he said in a conversation with AIM.

That gap inspired the creation of the Spatial-Temporal Graph Attention Network (STGAT). Unlike conventional LLMs that treat words as static relationships, STGAT introduces a time dimension. This matters especially in healthcare, where causality, the sequence of events, can mean the difference between an accurate diagnosis and a dangerous misdiagnosis.

Banerjee explains with an example: “A patient develops a rash a week after attending a pet gathering. A seasoned doctor immediately connects the dots. For AI, unless it understands the timeline, it’s just noise. STGAT builds that temporal understanding into the model.”

The result is a clinical knowledge graph already in use across India and in 200 clinics in Australia. And its scope extends far beyond clinical note-taking. Researchers are now exploring applications in drug discovery and clinical trials, essentially letting it act as an “intelligence layer” that integrates seamlessly with existing healthcare workflows.

Challenging the World on Speech Recognition

If reasoning is one frontier, voice is the other. Here too, the team’s work on Pingala V1, an automated speech recognition (ASR) system, has quietly made global history. The model has achieved a 2.94% word error rate for English and 3.1% for universal speech, outperforming heavyweights like NVIDIA, IBM and OpenAI’s Whisper on open leaderboards.

The name itself carries weight. The original inspiration behind the name, Pingala, was an ancient Indian sage, credited with inventing the binary representation of sound more than 2,000 years ago. This makes the model both a nod to Indic intellectual heritage and a declaration of intent. “We wanted to attribute the original creator of coding voice into binary,” Banerjee explained.

What makes Pingala’s achievement even more striking is its efficiency. Trained on just two GPUs in two days, it runs on commercial-grade hardware like NVIDIA’s L40. That means organisations can deploy world-class ASR for a fraction of the usual cost. Latency clocks in at under 100 milliseconds, a critical threshold for real-time applications such as telemedicine or multilingual customer support.

The model has been open-sourced on Hugging Face, where it has already been downloaded more than 2,000 times and licensed under a responsible AI framework that bars misuse. Pingala V1 supports 216 languages worldwide—including 39 Indian ones. The team estimates that it can understand 96% of the world’s population—something few Western players have even attempted.

The Case for Inclusive AI

The founders, many of whom come from small-town India, see firsthand how exclusionary design in AI could worsen social divides.

“If AI doesn’t understand someone speaking Santali in Jharkhand or Bhojpuri in Bihar, we’re designing the future to exclude them,” Banerjee warned. “It’s not about whether AI replaces humans. It’s about AI replacing people who don’t have access to it.”

To counter that risk, Shunya Labs has committed to releasing domain-specific ASR models for medicine and Indic languages. They also lean heavily on synthetic data generation and linguistic structure analysis, bypassing the scarcity of labelled datasets that has historically disadvantaged low-resource languages.

Privacy by Design

Unlike global platforms that centralise user data, Shunya Labs insists on on-premise deployment. Hospitals and enterprises can run their models locally, ensuring compliance with GDPR and India’s privacy norms without sending sensitive data to third parties. This, the team argues, is crucial for healthcare, where the stakes of data misuse are high.

Yet, for all their breakthroughs, the team is candid about the hurdles ahead. Data scarcity remains a constant challenge, as does a cultural scepticism towards Indian foundational research. “In the US, if you say you want to build a new model, the ecosystem rallies behind you. In India, the first question is: Why not use an American one?” Banerjee noted.

To change that, Shunya’s answer is a call for collective effort, from investors willing to back inclusivity, to media outlets highlighting open benchmarks, to government initiatives like Project Vaani that annotate neglected dialects. Equally insistent, they stressed, is academic rigour. “It’s not enough to claim success in PR. We need to publish, present at global conferences and invite public scrutiny,” he added.

In Sanskrit, ‘shunya’ means both zero and infinity, a fitting metaphor for what the company is attempting: zero word error rates, infinite possibilities. 

As Ritu Mehrotra, co-founder and CEO of Shunya Labs, puts it: “Small players can come, do real research, and put a model into the open domain, responsibly and for the world.”

Whether Shunya Labs will succeed in rewriting the global AI playbook remains to be seen. But with Pingala V1 already outperforming some of the world’s biggest names, one thing is clear: Shunya Labs is pulling the conversation on AI’s future to India.

The post The AI Race is About Scale. India is Asking if it Should Be appeared first on Analytics India Magazine.