Most people in AI circles know Dr. Viral B. Shah as one of the creators of the Julia programming language, a workhorse for scientific computing that now underpins everything from drug discovery to engineering simulation. Fewer realise that he also helped design Aadhaar, India’s biometric ID system, and the early plumbing of what became the India Stack, a suite of interoperable digital public infrastructure that today moves billions of real‑time payments and identity verifications every month. It is a rare combination: someone who has built both a modern programming language and a population‑scale digital identity and payments platform, and is now leading an AI‑native simulation company, JuliaHub, whose Dyad product brings agentic, physics‑grounded AI into the workflows of engineers building the physical world.
As governments everywhere scramble to “do something” with AI, often before getting their basic data, identity, and payments infrastructure in order, India’s experience looms large as both inspiration and cautionary tale. In this conversation, Shah argues that being AI‑ready is not about sprinkling machine learning on legacy systems, but about getting a handful of fundamentals right: clean, deduplicated identity; open, low‑cost interoperability; small, mission‑driven architecture teams with real executive backing; and a ruthless commitment to inclusion, from offline rural users to highly connected urban citizens.
Drawing on lessons from Aadhaar and India Stack, as well as his current work on scientific machine learning and digital twins at JuliaHub, Shah lays out the risks he refuses to take, such as deploying ungrounded, probabilistic AI on critical infrastructure, and the rules he believes every country should follow when building AI‑ready public systems. The result is a candid, globally relevant memo to heads of state, regulators, and public‑sector technologists about what it will take to make AI an asset, not a liability, in the next generation of public infrastructure.
Without clean, deduplicated core data, any layer you build on top, including AI, will suffer from a “garbage in, garbage out” problem.
When you think back to building Aadhaar and the payments rails that became the backbone of India’s digital economy, which design principle do you think is most transferable to any country that wants to build AI‑ready public infrastructure today?
Two foundational principles from Aadhaar are completely transferable to any country building AI-ready infrastructure today: strict uniqueness (deduplication) and open interoperability.
When we built Aadhaar, our first principle was creating a unique, deduplicated identification number for every resident. We did this by processing biometric data against the entire population to guarantee individual uniqueness. Without clean, deduplicated core data, any layer you build on top, including AI, will suffer from a “garbage in, garbage out” problem.
The second principle is building low-cost, scalable, and interoperable networks rather than closed, proprietary ones. For our payment infrastructure, we focused on a simple, universal set of just four core transactions: checking a balance, withdrawing money, sending money, and receiving money. Crucially, we prioritised interoperability so that no bank could serve only its own captive network of customers.
Because the Aadhaar ID was built to be inherently Know Your Customer (KYC)-compliant, any bank could instantly serve any Aadhaar holder without redundant verification. For AI-ready infrastructure, governments must similarly avoid building data silos. Systems must speak a common language so data can securely flow and train models effectively.
ALSO READ: ‘Do Not Fall in Love with the Prompt’: Ellis Crosby on Building Outcome‑First AI
When dealing with an enterprise as large and diverse as an entire country, engineers cannot just think about function; they must think about form, empathy, and intent.
You’ve written in Rebooting India about “small, focused teams of highly skilled people” being able to radically reimagine government systems. In the AI era, what kind of team should a head of state assemble if they want to build credible AI‑ready digital infrastructure in five years, not fifteen?
To build massive, nation-scale digital infrastructure, heads of state must resist the urge to build massive bureaucracies. They need to adhere to the famous “two-pizza team rule”: a small autonomous working group that is small enough to be fed by just two pizzas. This essentially helps keep the core architecture teams small, nimble, and highly cohesive.
The underlying technology and governance framework for Aadhaar, which now serves well over a billion people, was designed by a core team of approximately 10 to 15 people. Small teams ensure rapid communication, tight cultural bonding, and fast decision-making cycles.
In the AI era, technical ability is merely the baseline. The real differentiator for a sovereign AI team is a strategic, intersectional mindset. When dealing with an enterprise as large and diverse as an entire country, engineers cannot just think about function; they must think about form, empathy, and intent.
At JuliaHub, this is exactly what we focus on. By using advanced AI capabilities to automate repetitive, highly technical coding and optimisation tasks, we free up engineers to spend their time thinking about the actual intent and societal impact of the technology they are building. A great modern infrastructure team uses AI to accelerate its own workflow, ensuring the final product is genuinely fit for purpose across wide-ranging, diverse populations. We even built that philosophy into our products, like Dyad, that allows customers to take a broader view of the products they are building.
The real-world consequences of infrastructure failure are catastrophic—a collapsed bridge, a compromised power grid, or a broken welfare distribution system.
What’s a concrete example of a risk you’d refuse to take with AI in public infrastructure, even if the short‑term efficiency gains looked attractive?
In public infrastructure, you can never trade mathematical certainty for short-term efficiency. The real-world consequences of infrastructure failure are catastrophic—a collapsed bridge, a compromised power grid, or a broken welfare distribution system.
During the rollout of Aadhaar, we faced a concrete risk with biometric accuracy. We refused to rely solely on fingerprint scanning because the data showed it wasn’t accurate enough on its own to handle the “1-to-N” deduplication of a billion people. Relying on it blindly would have led to massive mismatches and systemic exclusion. To mitigate this risk, we insisted on multi-factor biometrics, introducing iris scanning alongside fingerprints to guarantee absolute uniqueness.
When we look at AI today, particularly Large Language Models (LLMs), they operate on probabilities and are prone to hallucinations. They do not understand the laws of physics or hard mathematical bounds. I would absolutely refuse to deploy ungrounded, probabilistic AI models to design or manage critical physical or civic infrastructure.
If you are using AI in infrastructure, it must be grounded in mathematical and physical certainty. This is why we advocate for Scientific Machine Learning (SciML), where AI models are strictly constrained by proven equations and rules that engineers use regularly. If an AI’s outputs cannot be tied to verifiable, deterministic rules, it is far too risky for public infrastructure. This is one of our fundamental architecture principles for our Dyad product from JuliaHub today. We ensure that laws of physics are grounding design recommendations from AI, based on reality. That gives our customers ultimate confidence in the results they get from AI solutions.
AI systems, unlike traditional databases, can make inferences and decisions that are hard to audit. How should governments think about auditability and explainability when they start putting AI in the loop of welfare delivery, credit schemes, or public services?
When AI impacts human lives, whether through welfare delivery, credit scoring, or public service access, black-box algorithms are entirely unacceptable. Governments must maintain rigorous, transparent audit processes to guarantee accountability. There are two ways to solve this.
First, on the technical side, if you are using AI for physical systems or resource allocation, the models should ideally be built using framework architectures like SciML, where the outputs are directly traceable back to hard equations, logical parameters, and verifiable data inputs. This makes the underlying “why” behind an AI’s decision auditable. Second, for socio-economic systems, AI should not be used as an unchecked decision-maker, but rather as an assistant for optimisation and anomaly detection.
ALSO READ: Dr Lee Schlenker’s Playbook For Boards Before EU AI Act’s Enforcement
For example, just as we used biometric matching to power seamless service delivery, governments can deploy AI to build intelligent fraud detection systems. Essentially, these systems should cut both ways: identifying individuals attempting to game the system, but also auditing the government itself to pinpoint exactly where it is failing to deliver critical services, like ration or fertiliser distribution. Ultimately, a human expert must always remain in the loop. AI can streamline workflows and flag discrepancies, but final accountability must rest on a clear, human-led audit trail.
When designing public infrastructure, you must factor in structural inequalities.
If asked about a “minimum viable stack” to be AI‑ready in ten years, what are the three building blocks you’d tell a government to prioritise first, and why?
Conversely, what are the most common traps you see governments falling into when they talk about AI and digital public infrastructure—things that sound modern but actually lock them into brittle systems?
If a government wants to be genuinely AI-ready over the next decade, they need to prioritise three foundational pillars:
- AI-Native Education at Scale: Governments need to re-engineer education to integrate AI-native curricula across all professional sectors, not just computer science, but engineering, medicine, and accounting, to build a modern workforce that knows how to build with and audit these tools.
- Sovereign Infrastructure and Entrepreneurship Ecosystems: Governments must encourage and foster a vibrant domestic ecosystem to support an AI-ready nation by ensuring that the pre-requisites are built, be it data centres, compute, or clean energy grids. Both infrastructure and policies need to be AI-ready to give local innovators and entrepreneurs the tools and ecosystem to innovate.
- Open, Anonymised Public Data Ecosystems: AI requires high-quality data. Governments sit on vast repositories of information. By securely anonymising and opening up large government datasets to the public, they create a rich training ground for building high-utility, localised models.
The Common Trap: The biggest trap I see governments falling into is chasing “modern-sounding” technology trends without considering basic accessibility and inclusion. They build brittle, hyper-advanced systems that look great on paper but ignore the ground reality of their population.
When designing public infrastructure, you must factor in structural inequalities. For instance, if you build an AI system that assumes every citizen has a smartphone and a high-speed internet connection, you are actively automating exclusion. In India, for example, roughly 65% of the population lives in rural areas, often using feature phones, and technology must be designed to address their needs.
ALSO READ: Middle East: The Sovereign AI Testbed US, EU and Asia Can Learn From
With Aadhaar, we designed the stack so that someone in a remote village, without a smartphone or a bank card, could still walk up to a local micro-ATM, authenticate via biometrics, and access financial services without traveling to a city branch. True AI-ready digital public infrastructure must empower the entire population, factoring in those with limited digital access from day one.
If you were to write a short memo titled “Five rules for AI‑ready public infrastructure in any country,” what rules would absolutely make that list?
There are a number of guidelines I could recommend, but the five most important rules I would identify are:
- Mandate Absolute Executive Leadership. Large-scale digital public infrastructure cannot be siloed within a minor ministry. It requires explicit, unyielding buy-in and driving force from the highest levels of government to cut through bureaucratic inertia.
- Build Small, Mission-Driven Architecture Teams. Adhere to a strict “two-pizza rule.” Keep your core design and implementation teams small (10 to 15 people), highly technical, and agile. Massive committees build slow, brittle systems; small teams build focused, scalable ones.
- Enforce Open Interoperability over Monopolies. Ensure all digital platforms share data seamlessly via open APIs. Never allow a single supplier, agency, or bank to capture a public network. True digital utility relies on low-cost, open rails where any verified node can communicate with any other.
- Implement Risk-Based, Enabling Regulations. Don’t let outdated legal frameworks choke innovation. Collaborate early with regulatory and financial authorities to create risk-based KYC and identity parameters that lower the barriers to entry for vulnerable populations, backed by short, sharp enabling notifications to activate digital laws.
- Ground Infrastructure AI in Certainty and Inclusion. Never deploy probabilistic, hallucination-prone AI systems where physical or civic safety is at stake. Ensure all AI infrastructure models are mathematically auditable, and explicitly design the stack to serve the offline, rural, and less-connected segments of your population from day one.
ALSO READ: From Siemens Energy to Bank of America: What “Quietly Advanced” Enterprises are Doing Differently
