Anupam Chattopadhyay, College of Computing & Data Science, NTU Singapore
Babak Hodjat, Cognizant
Balaji Thiagarajan, Flipkart
Hari Menon, Gates Foundation
Mike Haley, Autodesk
Raju Vegesna, Sify Technologies Ltd
Shereen Bhan, CNBC-TV18
SESSION OVERVIEW
For India and the Global South, as AI systems move from pilots to mission-critical deployment across enterprises and governments, this challenge is amplified by fragmented and heterogeneous data, legacy IT infrastructure, uneven regulatory maturity, multilingual and culturally diverse populations, and high-impact use cases in finance, healthcare, welfare delivery, and public services. This session will examine what enterprise-grade responsible AI looks like in such contexts, grounded in real deployments rather than abstract frameworks.
VIDEO RECORDING
Enterprise-Grade Responsible AI for India and the Global South
Detailed Summary
Sunita welcomed the audience, acknowledged the logistical challenges of attending the summit in Delhi, and framed the session as the closing discussion of the AI Impact Summit.
She highlighted the agenda: how to build guard‑rails, trust frameworks, and sovereign AI stacks for India and the Global South while ensuring safety, fairness, and scalability.
Balance of Trust – AI must avoid two extremes: over‑trust (treating AI as “magic pixie dust”) and over‑skepticism (requiring a human rubber‑stamp at every step).
Human‑in‑the‑Loop (HITL) & Agent‑in‑the‑Loop – Robust systems combine human oversight with automated agents that cross‑check each other’s outputs. Uncertainty estimation is used to decide when to route a decision to a human.
Error‑Correction & Redundancy – Analogous to telecom systems where bit‑flips are corrected through redundancy, AI pipelines should embed similar error‑detecting mechanisms.
Agentic Identity – As ecosystems incorporate third‑party agents (B2C, partner APIs), verifying an agent’s provenance becomes critical. Current standards are nascent; Google’s A2A work is cited as a leading effort.
Regulatory Tightrope – India must avoid both over‑regulation (stifling innovation) and under‑regulation (exposing citizens to unsafe systems). Sovereign LLMs are suggested as a pathway to maintain control while fostering local innovation.
Transition to Panel – Babak handed over after a brief recap, noting the difficulty of his role and the importance of “human‑centered” guard‑rails.
3. Research Challenges in the Global South (Anupam Chattopadhyay – NTU Singapore)
Challenges Highlighted
Heterogeneous Data & Intermittent Compute – Real‑world data in the Global South is noisy, multilingual, and often collected on low‑spec hardware.
Deep‑Fake Detection Case Study – Models trained on clean, high‑quality data perform poorly on noisy audio/video from Indian contexts; synthetic‑noise augmentation is required.
Synthetic Data Generation – Adding tunable noise and scraping large‑scale web data helps create robust training sets when real data is scarce.
Fact‑Checking Pipeline – For deep‑fakes, an automatic fact‑checker cross‑references images with trusted news sources to label content as “original” or “refake”.
Model Compression & Mixture‑of‑Experts – Deployments on limited hardware use a mixture‑of‑experts approach, where domain‑specific sub‑models handle particular tasks.
Federated Learning & Privacy – When organizations wish to merge proprietary models without exposing data, federated learning techniques preserve privacy while enabling collaborative improvement.
Call for Academia‑Industry Partnerships – Anupam advocated for a “single‑window” consortium (AI.SG) that links research funding, technology transfer, commercialization, dissemination, and regulation.
4. Sustainable AI Infrastructure (Mike Haley – Autodesk)
Infrastructure Pillars
Design‑First Sustainability – Data‑center design determines energy efficiency; sustainable design is the foundation of responsible AI.
Liquid‑Cooling Technologies – Provide high heat‑removal capacity with lower power consumption, enabling rapid scaling of AI workloads.
KPIs per Token – Proposes metrics such as energy‑per‑token and water‑per‑token to quantify environmental impact of AI inference.
Policy Recommendations – Governments should mandate KPI reporting and incentivize operators who meet or exceed sustainability thresholds.
Alignment with Global Discussions – Haley referenced Davos conversations on ROI and energy efficiency, emphasizing the need for renewable‑energy‑backed AI fleets.
5. Enterprise Perspective: Responsible AI at Flipkart (Balaji Thiagarajan – Flipkart)
Data Quality & Access Controls – High‑quality, well‑governed data is the prerequisite for trustworthy models; strict access‑control and encryption protect data at rest and in motion.
Domain‑Specific Models (SLMs) – Flipkart creates small, region‑specific language models (e.g., for Mumbai vs. Delhi pricing) to ensure localized fairness.
Image‑to‑Catalog Pipeline – An AI system converts seller‑uploaded product images into structured catalog listings within minutes, using domain‑specific vision models.
Transparency in Bot Interaction – Customer‑service agents operate as co‑pilots; a disclaimer informs users they may be interacting with a machine, with an opt‑out default to protect trust.
Mixture‑of‑Experts Orchestration – A dynamic agentic framework selects between large LLMs (for intent detection) and SLMs (for fine‑grained personalization).
Future Roadmap – Continuous refinement of the orchestration layer, expansion of synthetic‑data pipelines, and tighter compliance with emerging regulations.
6. Government AI Stack Recommendations (Babak Hodjat – second contribution)
Suggested Framework
Public Processing Capacity – Create a publicly accessible compute pool to democratize access for startups, academia, and public‑sector innovators.
Sovereign Sandbox – Establish a controlled environment where regulators, startups, and researchers can trial agentic systems and iterate on policy without systemic risk.
Ecosystem‑Centric Role – Government should nurture an ecosystem rather than build proprietary stacks; enable talent attraction, open‑source contributions, and transparent standards.
Balanced Regulation – Avoid front‑running regulation while preventing negligent laissez‑faire; sandbox‑driven policy evolution is the recommended path.
Alignment with National Initiatives – References India’s AI‑mission GPUs allocation to states and institutions, and the emergence of open‑source sovereign LLM initiatives.
Collective Sentiment – Panelists expressed optimism about India’s capacity to leapfrog traditional SaaS models into AI‑first products, citing the nation’s massive talent pool, cost advantage, and governmental support.
Sutra of “People, Planet, Progress” – Emphasized that responsible AI must serve citizens, protect the environment, and drive sustainable economic growth.
Audience Interaction – Several informal interjections (e.g., requests for a group photo, queries about AI usage in summit logistics) highlighted the energetic atmosphere.
Final Call to Action – Participants were urged to disseminate the “Sarvajana Hittaye, Sarvajana Sukhaye” vision (welfare and happiness for all) beyond India, to the Global South and to international partners.
Key Takeaways
Balanced Trust is Essential – Over‑trust and over‑skepticism both erode AI reliability; layered human‑and‑agent oversight with uncertainty quantification offers a pragmatic middle ground.
Data Diversity Drives Model Robustness – Synthetic‑noise augmentation, federated learning, and domain‑specific models are critical for handling heterogeneous, multilingual data typical of the Global South.
Sustainable Infrastructure Must be Designed First – Liquid‑cooling, renewable‑energy KPIs (energy‑per‑token, water‑per‑token) and modular data‑center architecture are foundational to responsible AI at scale.
Transparency in Human‑Machine Interaction – Explicit bot disclosures with an opt‑out default safeguard user trust, especially in high‑volume consumer‑service settings like Flipkart.
Public Compute Resources & Sandbox Regulation – Government‑provided processing capacity and a sovereign sandbox enable inclusive innovation while allowing regulators to test policies safely.
Domain‑Specific “Small” Models (SLMs) Complement LLMs – A dynamic orchestration layer that routes queries to either large foundational models or narrowly‑tuned small models improves relevance, fairness, and resource efficiency.
India’s Position as an AI Leapfrog Nation – Leveraging its service‑industry heritage, massive user base, and proactive policy framework, India can become a leading provider of sovereign AI solutions for the Global South.
Responsible AI Must Align with “People, Planet, Progress” – Ethical, environmental, and socioeconomic dimensions are inseparable; any AI deployment should be evaluated against this three‑fold mantra.
Collaborative Academia‑Industry Consortia Accelerate Progress – Initiatives like AI.SG demonstrate how coordinated funding, technology transfer, and regulation can close the gap between research and enterprise deployment.
Continuous Learning & Adaptation – The panel emphasized that responsible AI is a moving target; ongoing monitoring, feedback loops, and policy updates are required to keep pace with rapid AI advances.