Best Place To BuildPratyush Kumar, Co-founder, Sarvam AI | "Sarvam means everybody- AI should be for everyone."| Ep. 24
CHAPTERS
Why Indian-language AI matters: diversity, culture, and strategic tech
Pratyush frames the core problem Sarvam/AI4Bharat set out to solve: building AI that truly works across India’s linguistic and cultural diversity. He argues that for strategic technologies like AI, the country should retain the capability to build key systems itself.
- •India’s many scripts and language variations make Indian-language AI a uniquely hard challenge
- •Culture is embedded in language; AI must reflect both linguistic and cultural context
- •Strategic technologies require domestic capability, not just adoption
- •Building locally changes a country’s stance and autonomy
Pratyush’s path into AI: systems engineering to deep learning scale
He traces his journey from electrical engineering and systems/HPC research into AI during the early deep-learning inflection point. The discussion highlights the shift from algorithm-centric progress to compute-and-data-driven scaling and why that pulled him into foundational work.
- •Education: IIT Bombay (EE) → ETH Zurich PhD (systems/HPC) → IBM Research
- •Deep learning’s early era (post-AlexNet) and the move toward compute-driven AI
- •Systems reliability, performance, and scale as crucial AI enablers
- •Joining IIT Madras faculty and deciding to focus on AI research for India
AI4Bharat origins: teaching at scale, community experiments, and a pivot to language
AI4Bharat begins with hands-on deep learning courses that unexpectedly scale to tens of thousands of learners. After a decentralized ‘hacker motion’ doesn’t work well, the team focuses the effort on a few lab-led problems—eventually converging on Indian-language AI as the highest-leverage direction.
- •Hands-on deep learning courses; ~50,000 learners and real career impact
- •Initial decentralized volunteering model didn’t sustain execution
- •Shift to lab-driven problems with students + volunteers
- •Strategic pivot: Indian-language AI becomes the central mission
Building foundational components at IIT: data pipelines to competitive translation
Pratyush explains how students helped build critical building blocks such as high-quality web scraping for Indian-language data. Within about a year, the lab produced translation systems competitive with big tech, attracting government and philanthropic support and maturing into a large center of excellence.
- •Early ‘unsexy’ foundations: high-quality scrapers and data flywheel setup
- •Rapid progress to competitive English↔Indic translation systems
- •Support: Government of India (KASHINI) + Nandan Nilekani philanthropy
- •AI4Bharat scales to ~200+ people; leading open-source Indic language lab
- •Ongoing collaboration between Sarvam and AI4Bharat
From lab to company: why Sarvam AI was created
The conversation shifts to why a venture-backed company was needed: foundational models and production-grade LLMs require far more compute, capital, and engineering than translation systems. Sarvam is positioned as a contrarian bet on India’s market potential and long-term AI cycle, inspired by DPI scale stories like Aadhaar/UPI.
- •LLMs require far more compute/capital than earlier NLP systems
- •Motivation: build ‘from scratch’ in India, for India—beyond short-term apps
- •Co-founder Vivek Raghavan’s background: Aadhaar, DPI, and applied AI systems
- •Contrarian view: India isn’t “small market”; it’s early in a long AI adoption cycle
- •Public-private platforms (UPI-like) as a model for scaling AI
What foundation models are—and why nations and companies race to build them
Pratyush defines foundation models as general-purpose systems that can be adapted to many tasks rather than single-skill models. He connects their importance to broad economic value creation and to AI’s emerging status as strategic national infrastructure.
- •Foundation models: general-purpose capabilities across language/audio/vision
- •Applications sit ‘on top’ (e.g., search, code assistants, legal summarization)
- •Rapid cycles: months instead of decades compared to prior industrial revolutions
- •Commercial imperative plus strategic/national-security imperative
- •OpenAI/Google/Meta/China as examples of global-scale builders
Sarvam’s four-layer ‘full-stack’ view: inference, models, orchestration, applications
He lays out Sarvam’s stack and why the company spans layers instead of specializing in one. The orchestration layer is emphasized as the glue that makes real systems reliable—especially for voice and reasoning-heavy workflows—while domain experts drive application design.
- •Inference layer: efficient serving, GPU utilization, compilation/optimization, cost control
- •Model layer: LLMs and other modality models (audio/vision)
- •Orchestration layer: combining models + tools + software for reliable, scalable systems
- •Applications layer: domain-led solutions (e.g., insurance, government services)
- •Rationale: different skills needed at each layer; full-stack helps deliver outcomes
Data for India: scarcity, ‘culture tokens,’ and code-mixed realities
Pratyush details why Indic data is harder: many languages have limited digitized text and cultural content often exists in undigitized sources. He also highlights code-mixing and Romanized typing as essential real-world phenomena models must handle to be useful for Indian users.
- •Three core ingredients: data, compute, talent (plus non-obvious know-how)
- •Indic languages: uneven data availability and incomplete digitization
- •‘Culture tokens’: content embedded in old/undigitized books and artifacts
- •Modern usage: Romanized Indic typing and evolving language norms
- •Code-mixing: English words within Indic sentences as a first-class requirement
Real deployments in India: Aadhaar basement stacks, insurance calls, courts, and NITI Aayog
Concrete case studies illustrate what ‘full-stack’ means in practice: air-gapped sovereign deployments, large-scale voice outreach, and complex data-to-policy reasoning systems. The examples stress reliability, latency, and application-specific design as key to real value creation.
- •Air-gapped, on-prem deployments for UIDAI/Aadhaar-style constraints
- •Voice bots assisting citizens when biometric authentication fails
- •Insurance renewal outreach at crore-scale in local languages
- •Judgment accessibility: working with courts; legal language isn’t plain translation
- •NITI Aayog project: querying thousands of tables + PDFs via natural language; auto-code, self-correction, report generation
Strategic autonomy and sovereign AI: capability to build, deploy, and scale
Pratyush frames sovereign AI as the ability to build strategic technology domestically without ‘decoupling’ from the world. He links the concept to national resilience, large-scale deployment capacity (compute/power/app ecosystem), and staying close to state-of-the-art through a long AI cycle.
- •Strategic autonomy: collaborate globally, but retain ability to build from scratch
- •Sovereign AI includes talent + resources + deployment capability at national scale
- •AI as a general-purpose technology akin to electricity in development impact
- •Goal: start now to stay competitive through 2030 and beyond
- •Global momentum: Japan, South Korea, France, UK and others launching programs
AI as a utility: open standards, per-capita access, and technology as an equalizer
The discussion explores how AI could mirror India’s digital public infrastructure pattern: utility-like access with open standards that enables private innovation. Pratyush suggests ‘per-capita AI consumption’ as a future proxy for competitiveness and argues that AI can flatten access gaps between urban and rural users.
- •AI should be cheap, efficient, and widely available like UPI-style utilities
- •Open standards vs open source: the right model is still evolving for AI
- •Vision: citizens get access to ‘lakhs of tokens per day’ via useful applications
- •Per-capita AI usage as a proxy for national productivity/advancement
- •Equalized experience: rural and urban users should get comparable service quality
The economics of building AI: cost drivers, why funding matters, and the local value loop
Pratyush breaks down where money goes: data preparation, GPU-heavy training, expensive talent, and productization across the stack. He introduces the ‘value loop’—deploy, learn from usage, improve quickly—and argues that keeping this loop within India is key for long-term economic and strategic returns.
- •Major costs: data prep, model training, GPUs, and top-tier talent
- •‘GPU factories’ idea: GPUs also process/produce training datasets
- •Beyond models: software engineering + orchestration + applications add cost
- •Costs are falling with efficiency improvements, but still prohibitive at scale
- •Value loop: deployments create feedback → better models in months; must remain local
Competition speed, focus, and building amid constant change
The conversation turns to the emotional and operational reality of the AI race: rapid shifts, noisy news cycles, and strong capital flows. Pratyush emphasizes anchoring on medium-term clarity and democratization goals to avoid being whipsawed by short-term hype.
- •AI progress is ‘too fast for comfort’ and can feel overwhelming
- •Need strong medium-term vision to stay steady amid noise
- •Maintain focus despite abundant capital and constant new announcements
- •Sarvam’s north star: democratizing AI for India over decades
- •Plan for compounding gains leading up to 2030
Ecosystem and talent pipeline: academia-startup integration and a ‘builders vs sellers’ mindset
Pratyush describes the early signs of an ecosystem effect around IIT Madras but says deeper integration is needed among academia, VCs, and operators. He argues for hands-on building, real feedback loops, and structures that help students (beyond IITM) execute in a world changing every six months.
- •Ecosystem exists but needs stronger institutional bridges between campus and startups/VCs
- •Encourage students’ self-belief and exposure to real building environments
- •‘Cut the crap’: prioritize building with fast feedback over abstractions
- •Open doors beyond IITM; leverage research park and broader participation
- •National deep-tech moment: more government and strategic sector engagement
AI in daily workflows, hallucinations, and the human–AI relationship
They discuss how AI is reshaping everyday work and the risks of over-reliance, including stacked hallucinations in research workflows. The conversation broadens into philosophy: what remains uniquely human, how engineered systems shape experience, and the need to deliberately steer toward a positive, symbiotic future.
- •AI now sits ‘in the palm’—directly used across teams and weekly workflows
- •Reliance risks: hallucinations and cascading errors across toolchains
- •Concern: future gap between human and machine capability
- •Reframing humanity beyond output; prioritizing lived experience and values
- •Engineered systems matter—society must deliberately design for positive outcomes
Sarvam’s roadmap: sovereign LLM build, scaling products, closing the state-of-the-art gap
Pratyush closes with Sarvam’s near-term execution plan and multi-year ambition. The company aims to ship India’s sovereign model, expand product surface area, and steadily reduce the gap to global state-of-the-art while keeping democratization at the center.
- •Near term: build the sovereign model within the year
- •Current products already scaled (millions of calls; analytics for organizations)
- •Sovereign model expands ‘surface area’ for new applications and partners
- •2–5 year goal: close gap to best models of the day as the frontier shifts
- •Long-term intent: strategic capability + broad access for India’s population scale
