Webinar: Scalable & Reliable Data Systems: The Foundation for Enterprise AI
For data architects, platform engineers, and technical leaders bringing AI into mission-critical systems.
Key Takeaways
- Design for failure from day one. Reliability, availability, and scalability can't be engineered in later — they have to be part of the architecture from the very first design decision, because every component will eventually fail. The real question is how fast your system detects the failure and how it behaves while recovering. (Tony Fernandes)
- Architect for 10x–100x your current scale, not just today's requirements. Building for tomorrow's load from the outset — through stateless services, horizontal scaling, and asynchronous processing — avoids a full redesign later. (Tony Fernandes)
- In payments and other mission-critical systems, correctness matters more than speed. A slightly delayed transaction is inconvenient; an incorrect one destroys trust. That means real investment in consistency guarantees, idempotency, reconciliation, and auditability — a different bar than most consumer applications need to meet. (Tony Fernandes)
- True active-active architecture is far harder than it looks — and rarer than vendors imply. Building it requires thinking in CAP-theorem terms (is a given system, or part of it, CP or AP?), and — critically — it requires deep involvement from application developers, not just infrastructure teams, to classify which reads and writes are failure-tolerant versus mission-critical, and to design real conflict-resolution policies for multi-region writes. (Tony Fernandes & Amalendu Chakraborthy)
- Security and privacy have to be architected in from the ground up, not bolted on. Apple's Secure Enclave is a concrete example: encryption keys are generated per-device and never leave the chip — not even Apple can access them, which is why the company was able to decline government requests to unlock a specific iPhone. (Amalendu Chakraborthy)
- Analytics and transaction processing are converging into a single real-time pathway. The old model — process the transaction fast, run analytics offline afterward — is disappearing. Fraud detection, personalization, and other "intelligence" work increasingly needs to happen in the transaction path itself, not after it. (Tony Fernandes)
- LLM inference speed is fundamentally mismatched with how computing infrastructure was designed. Traditional systems operate on a scale from nanoseconds (cache) to milliseconds (network); LLM inference, which takes full seconds, doesn't fit that paradigm at all — which is why next-generation compute architecture is being rebuilt around making AI inference a first-class, cache-coherent part of the CPU/memory complex, not an offline accelerator bolted on the side. (Sajjad Khazipura)
- Unified multi-model databases (one engine for graph, vector, and relational data) show real promise but remain unproven at true scale. The underlying principle — that it's the same data, just consumed through different "shapes" depending on the query — has merit, but horizontal scalability for graph and vector workloads specifically is still maturing across the industry. (Amalendu Chakraborthy & Sajjad Khazipura)
- Put an abstraction layer between your application and your database choice. Today's best database decision may not be the right one in two years — an abstraction layer means swapping technology later doesn't require tearing apart your application. (Tony Fernandes)
Q&A Highlights
The following questions were raised by the panel moderator during the live session.
What's the biggest challenge in bringing AI into existing legacy systems?
Less technical than organizational. In a large enterprise, any significant architecture or application change touches many stakeholders who each have a say — and building consensus for that change takes real, sustained effort. It's fundamentally a change-management problem, not a technology problem. (Amalendu Chakraborthy)
There's also a distinction worth drawing between two different uses of AI:
Using it to make engineering teams more productive (increasingly uncontroversial — most companies have already adopted this) versus embedding AI directly into a product's core functionality, which is still being actively figured out, and looks different for every company depending on what they actually build. (Tony Fernandes)
If you were architecting a new company's data systems today, would you do anything differently to accommodate AI?
The industry now has considerably more mature tooling — vector databases and graph databases in particular — that weren't practical options even a few years ago. That said, horizontal scalability remains a real limitation for both graph and vector databases today; the industry is actively working on this, but it isn't solved yet. The core architectural principle doesn't change: your data architecture still needs to be linearly scalable, and you should design with an awareness of which newer data structures can genuinely help your specific AI workloads. (Amalendu Chakraborthy)
Should enterprises consolidate around one unified database, or use specialized databases for different data types (graph, vector, relational)?
Vendors like Oracle are increasingly promoting a single database engine that natively supports relational, graph, vector, and document data together — but whether that actually delivers at true production scale is still an open question the industry hasn't fully answered. There's real merit to the underlying idea, though: it's ultimately the same data, and the "shape" you query it in — like a stencil placed over a light source — matters less than getting the underlying data model right in the first place. (Amalendu Chakraborthy & Sajjad Khazipura)
Is AI ready to autonomously manage and support production systems, not just help build them?
It's a matter of time and building justified confidence — the same pattern the industry went through with earlier waves of automation. Historically, teams wanted to observe many consistent instances of the same signal before trusting an automated system to take corrective action on its own; the open question with AI is exactly the same: how sure can you be that the model's corrective action won't make things worse? Some companies are already starting down this path, but broad adoption will take time. (Tony Fernandes)
Panelists
Tony Fernandes
Formerly Vice President, Payment Processing Architecture, Visa
Tony is a seasoned technology executive with over three decades of experience leading enterprise technology strategy, digital transformation, and architecting large-scale program delivery. He has held senior leadership roles at global organizations, including most recently at Visa Inc., where he has driven the architecture of complex, mission-critical initiatives at scale. Tony is known for his ability to bridge business and technology, aligning engineering execution with strategic objectives to deliver measurable outcomes. Throughout his career, he has led cross-functional teams, modernized legacy platforms, and implemented innovative solutions in highly regulated, high-availability environments.
Amalendu Chakraborty
DaaX Advisor; formerly at Apple & Oracle
Amalendu is a senior technology leader and systems architect with over two decades of experience. He is passionate about applying AI and advanced data platforms to drive meaningful business transformation. He has successfully built and scaled global data systems across multi-country enterprises, delivering resilient, high-performance architectures that support mission-critical operations. He brings a unique ability to translate emerging technologies into a durable business advantage while leading high-impact teams at scale.
Sajjad Khazipura
Co-Founder & CTO, DaaX
Co-Founder & CTO of DaaX. At DaaX, Sajjad is building LAKEer — an enterprise neuro-symbolic AI platform combining knowledge graphs, vector search, and structured data retrieval to deliver grounded, hallucination-resistant AI for mission-critical applications. He brings experience building internet-scale, high-reliability systems, petascale data platforms, real-time transactional systems, and industry AI solutions across e-commerce, manufacturing, healthcare, and finance. Prior roles include Principal Solutions Architect at AWS and VP & Global Practice Head at Wipro Technologies.
Next webinar
- Webinar: Introduction to Neuro-Symbolic AI
- Continue the DaaX webinar series with a practical introduction to Neuro-Symbolic AI.