Best Place To BuildPratyush Kumar, Co-founder, Sarvam AI | "Sarvam means everybody- AI should be for everyone."| Ep. 24
CHAPTERS
- 0:00 – 2:34
Why India needs AI for its languages (and why it’s strategic)
Pratyush frames the core motivation: India’s linguistic diversity makes Indian-language AI both a hard technical challenge and a cultural necessity. He argues that strategic technologies like AI require the capability to build domestically, not just consume from abroad.
- •India’s diversity shows up most sharply in scripts, dialects, and speech patterns
- •Building Indian-language AI is a uniquely challenging and valuable problem
- •Domestic capability changes a country’s strategic posture
- •AI should be treated as a strategic technology for India
- •Sets the tone for “AI for everyone” as a national + societal goal
- 2:34 – 4:11
Pratyush’s path: systems engineering to deep learning at scale
He traces his journey from electrical engineering and high-performance computing to early deep learning work at IBM Research. The key shift he highlights is AI moving from algorithm-centric breakthroughs to compute- and data-driven scaling.
- •EE background built foundations in math + coding
- •PhD at ETH Zurich focused on reliable/high-performance systems
- •At IBM Research, saw the early deep learning shift post-AlexNet
- •AI progress increasingly driven by compute, data, and systems efficiency
- •Decision to pursue fundamental AI research in India
- 4:11 – 6:29
AI4Bharat begins: from teaching deep learning to building for India
After joining IIT Madras, Pratyush and Prof. Mitesh Khapra start by building a large hands-on deep learning community. That community momentum evolves into AI4Bharat, aimed at applying best-in-class methods to India-relevant problems.
- •Joined IIT Madras and partnered with Mitesh Khapra (around 2017)
- •Started with large, hands-on deep learning courses
- •Built a broad community via courses, hackathons, and projects
- •Chose India-focused problems rather than generic benchmarks
- •Converged on Indian-language AI as the flagship direction
- 6:29 – 7:30
How students and open infrastructure built foundational components
Pratyush explains how AI4Bharat shifted from a decentralized volunteer model to a lab-led effort with students and volunteers. Early “unsexy but critical” work—like high-quality web scraping—enabled a data flywheel that unlocked competitive translation systems.
- •Pivoted to a tighter lab-driven execution model
- •CFI students were early builders; student work became core infrastructure
- •High-quality Indian-web scraping as a foundational capability
- •Within a year, built translation systems competitive with big tech
- •Momentum attracted funding and partnerships
- 7:30 – 8:31
AI4Bharat scales into a national-level center of excellence
AI4Bharat grows from a lab into a major open-source Indian-language AI effort, supported by government and philanthropy. It becomes a large interdisciplinary team and a key collaborator for Sarvam’s LLM ambitions.
- •Government of India support via KASHINI project
- •Philanthropic support including Nandan Nilekani/Nilekani Foundation
- •Now a large center (hundreds of people) at IIT Madras
- •Positions as a leading open-source Indian-language AI lab
- •Ongoing Sarvam–AI4Bharat collaboration for LLM-era work
- 8:31 – 12:25
Birth of Sarvam AI: going beyond translation to foundational models
Pratyush explains why a venture-backed company was needed: training production-grade LLMs requires substantially more compute and capital than earlier systems like translation. He introduces co-founder Vivek Raghavan and the contrarian thesis that India is a large, long-horizon market for foundational AI.
- •LLMs require far more compute/capital than earlier language systems
- •Vivek Raghavan’s background: startups + Aadhaar + DPI work
- •Mission: build LLMs and next-gen AI “in India, for India”
- •Contrarian bet against ‘small market/low ARPU’ narratives
- •Analogy to Aadhaar/UPI scaling beyond expectations
- 12:25 – 16:07
What a foundation model is—and why everyone is building them
He defines foundation models as general-purpose systems that become a base layer for many applications. The discussion covers why the world is racing to build them: they resemble general-purpose technologies like the steam engine, with rapid iteration cycles and both commercial and strategic value.
- •Foundation models are general-purpose, not task-specific
- •Used as a base to build many applications (search, legal, support bots)
- •Rapid cycles: months, not decades; digital replication accelerates spread
- •Commercial imperative + strategic national imperative both apply
- •Strategic tech parallel: space/nuclear-class importance
- 16:07 – 18:11
Sarvam’s four-layer ‘full stack’ view: inference → models → orchestration → apps
Sarvam positions itself as a full-stack AI company and breaks the stack into four layers. Pratyush clarifies why orchestration is distinct: real-world systems require combining models, tooling, latency/reliability engineering, and domain-specific design.
- •Inference: running models efficiently (cost, GPUs, compilation, optimization)
- •Models: LLMs + vision/audio models, potentially many per use case
- •Orchestration: combining models + code into scalable, reliable systems
- •Applications: built by domain experts without needing to manage lower layers
- •Two orchestration examples: real-time voice bots and complex reasoning systems
- 18:11 – 21:17
The hardest ingredient: Indian-language data (culture, code-mixing, Romanization)
He details the data challenge for Indian languages: multiple languages with uneven digitization, plus ‘culture tokens’ locked in undigitized materials. He also emphasizes modern language realities like Romanized typing and code-mixing, which models must learn to handle naturally.
- •Training needs enormous data: trillions of words; massive audio hours
- •Many Indian languages lack sufficient digitized data
- •Need both ‘language tokens’ (fluency) and ‘culture tokens’ (context/history)
- •Romanized input is a real usage mode and must be supported
- •Code-mixing (English words inside Indic sentences) is normal and important
- 21:17 – 27:17
Real deployments: Aadhaar-grade air-gapped systems, insurance outreach, courts, and policy analytics
Pratyush grounds the stack in concrete deployments, including running GPU boxes inside air-gapped Aadhaar infrastructure for citizen support calls. He also covers enterprise-scale voice outreach, legal translation/simplification for courts, and a NITI Aayog system that answers questions by selecting datasets, writing code, and generating reports.
- •Air-gapped UIDAI/Aadhaar deployments with on-prem GPU ‘boxes’
- •Voice bots for citizens (e.g., calls when biometrics fail)
- •Insurance renewals/education via multilingual voice at crore-scale
- •High courts: making judgments accessible in local languages (beyond naive translation)
- •NITI Aayog: natural-language Q&A over thousands of tables + PDFs with code-writing and self-correction
- 27:17 – 35:02
Sovereign AI and strategic autonomy: capability to build, scale, and stay competitive to 2030+
He defines sovereign AI as a country’s ability to build strategic AI systems from scratch while still collaborating globally. The goal is not only ownership but capacity: talent, compute, deployment at population scale, and a long-horizon plan to stay close to state-of-the-art through a multi-decade technological cycle.
- •Strategic autonomy: don’t decouple, but retain the ability to build end-to-end
- •AI framed as strategic tech akin to defense-relevant capabilities
- •Scale matters: compute, power, developers, deployment pathways
- •AI consumption could become a development proxy like electricity usage
- •Optimization target: where India stands vs state-of-the-art in 2030 and beyond
- 35:02 – 40:52
AI as a utility: India’s DPI template, openness, and per-capita ‘tokens’
Pratyush compares AI’s future in India to the public-private DPI model behind Aadhaar/UPI: open standards, government-trust catalysis, and private innovation on top. He suggests a future where citizens effectively get abundant ‘tokens’ per day through low-cost utility access, making AI a societal equalizer.
- •DPI lesson: open standards + government trust/scale + private innovation
- •AI world hasn’t finalized what should be open standard vs proprietary
- •Utility framing: make serving efficient and cheap to unlock broad innovation
- •Per-capita AI usage as a scale/competitiveness indicator
- •Equalizing effect: consistent experience across urban and rural users
- 40:52 – 45:11
Economics of building AI: GPU ‘factories,’ talent costs, and the local value loop
He breaks down why AI needs significant funding: data preparation, model training on large GPU clusters, and expensive top-tier talent, plus engineering and productization. He emphasizes the fast feedback ‘value loop’ where deployments quickly improve models—so keeping that loop local matters for India’s long-term advantage.
- •Primary costs: data processing, training runs, and scarce top talent
- •GPU clusters as ‘factories’ producing models and even processed datasets
- •Beyond models: orchestration + reliability + applications require major engineering
- •Costs are falling, but only ‘tens’ of serious foundation builders may exist
- •Value loop: deploy → learn → improve in months; keep the loop in-country
- 45:11 – 56:21
Operating in a fast-moving field: focus, ecosystem-building, and integrating academia with startups
Pratyush describes the pace as ‘too fast for comfort’ and argues that clarity of medium-term goals is crucial amid noise and hype. He calls for stronger campus-to-VC/startup linkages, a builder mindset, and institutional structures that accelerate real feedback over abstractions.
- •Need stable focus on medium-term horizon despite weekly/monthly shifts
- •Sarvam’s anchor: democratizing AI for India over decades
- •Ecosystem: AI4Bharat + Sarvam show the potential, but more integration is needed
- •Encourage ‘builders or sellers’—real feedback loops over abstractions
- •Bring experienced founders to help build structures, not just give talks
- 56:21 – 1:04:01
Living with AI: from daily workflows to human meaning, then Sarvam’s roadmap
The conversation shifts from practical workflow adoption (including LLM hallucinations) to philosophical questions about human identity and engineered systems. Pratyush closes by outlining Sarvam’s plan: build the sovereign model this year, scale products, and close the gap toward state-of-the-art over 2–5 years while keeping democratization central.
- •AI already reshaping workflows; risks include compounded hallucinations
- •Machines rapidly improve with unclear upper bounds; humans may become dependent
- •Human experience remains intrinsic, but systems must be engineered deliberately
- •Near-term: build India’s sovereign model; expand product surface area and scale
- •2–5 year goal: stay near state-of-the-art while democratizing AI access in India
