Thanks for being here today. I’m Pavan Gondhi, with over 20 years of experience steering through digital shifts as an executive, and I want to share my take on AI; how it’s truly unlocking data’s game-changing potential.

By 2030, AI might pump $15.7 trillion into the global economy, all fueled by data. But honestly, too many companies are buried in data overload, stuck with silos and old systems that block real insights. AI flips the script, think machine learning nailing customer predictions, or generative AI revamping workflows in biopharma, mining, and finance, slashing costs and downtime.

The future? Thrilling: Federated learning democratizes data for things like personalized healthcare, plus dashboards tracking AI’s real-time impact. But we’ve got hurdles—biases, ethics, cyber threats, and job shifts mean we need strong governance and training.

Here’s my advice: Audit your data now, test AI pilots with solid metrics, and rally your teams. Let’s seize this and build a stronger tomorrow.

Modern lakehouse platforms ingest data from transactional systems, event streams, APIs, and unstructured content, all of which evolve independently. Traditional ETL pipelines and schema registry based validation approaches struggle to keep pace with schema drift, semantic changes, and cross format variability, often requiring frequent manual intervention. Prior research shows that static schema validation captures only a narrow subset of data quality issues in large scale, continuously changing systems.

This session presents an AI governed multi modal data ingestion architecture that adapts autonomously to both structural and semantic change. The approach introduces semantic contracts that capture learned value distributions, attribute relationships, and temporal patterns derived from historical data, extending validation beyond basic syntactic checks. Building on established work in statistical validation and temporal modeling, the system enables real time semantic coherence assessment without rigid schema enforcement.

A self evolving schema intelligence layer uses large language models and embedding based similarity scoring to interpret schema evolution events, infer semantic equivalence across heterogeneous sources, and automatically generate transformation logic. These capabilities align with published research on few shot learning, automated schema matching, and contextual embeddings rather than custom heuristics.

The architecture unifies structured, semi structured, and unstructured inputs through AI driven extraction, recursive parsing, and event time alignment based on the Dataflow model. Apache Flink provides exactly once stream processing and state management, while Kubernetes supports elastic scaling, automated recovery, and zero downtime operator updates.

This talk is aimed at data scientists, platform architects, and data engineers building resilient ingestion systems for rapidly evolving lakehouse environments, with a focus on design decisions, tradeoffs, and lessons learned when integrating AI driven governance into distributed streaming pipelines.

The data problems that drive this are real:

• Gaps in training data — if the model never saw reliable information on a topic, it fills the void with plausible-sounding noise.
• Noise and contradictions in the corpus — the web is full of misinformation, outdated facts, and conflicting claims. Models absorb all of it.
• Memorization vs. generalization — models sometimes over-generalize patterns, producing confident-sounding responses that have no factual basis.

Unstructured operational directives often contain ambiguity, missing context, and hidden dependencies that make automation difficult. This session introduces a self-correcting agentic extraction framework built on a knowledge graph to automate trust operations workflows. The talk shows how schema-guided reasoning, structural and semantic validation, and iterative feedback loops enable agents to detect errors, reconcile inconsistencies, and refine outputs. Beyond the use case, the session highlights reusable design patterns for building reliable agentic systems for complex enterprise documents.

Abstract coming soon.

As artificial intelligence adoption accelerates across banking, model performance is still too often evaluated using narrow technical metrics such as accuracy, precision, and recall. While necessary, these measures fail to capture the true economic and risk impact of AI systems operating in highly regulated, judgment-driven environments such as fraud detection, AML monitoring, credit risk, and operational decisioning.

This session proposes a risk-adjusted KPI framework for measuring AI value in banking—one that moves beyond predictive accuracy to quantify how models influence risk outcomes, resource allocation, and regulatory exposure. We introduce practical metrics such as risk-weighted alert capture, marginal risk reduction per analyst hour, capital-at-risk efficiency, false positive cost of control, and human-in-the-loop uplift, linking model outputs directly to business and risk objectives.

The discussion concludes with guidance for model owners, risk leaders, and compliance teams on designing KPI frameworks that satisfy SR 11-7 model risk management, support regulatory transparency, and enable boards to assess AI investments through a risk-return lens—ensuring AI is measured not just by what it predicts, but by how it changes decisions and outcomes

Many organizations struggle with AI adoption—not because of technology, but due to lack of clarity, alignment, and readiness. This session shares a real-world approach to designing and implementing an AI strategy within a legacy enterprise, focusing on simplifying complexity and driving measurable impact. Learn how prioritizing data strategy, leveraging existing tools, and bridging business and technical teams can unlock value without heavy investment. The talk also highlights the importance of AI readiness across data, infrastructure, and culture—showing how strong foundations, not just models, determine success.

In today’s data-driven enterprises, making timely, accurate decisions requires transforming raw information into actionable insights at unprecedented speed. In “Real-Time at Scale,” I will explore how modern data platforms ingest, process, and act on streaming data, highlighting architectures and trade-offs that enable instantaneous decisions. Moving to “Analytics in the Age of LLMs,” I’ll demonstrate how generative AI and large language models are reshaping dashboards, metrics, and decision workflows, turning traditional reporting into dynamic, intelligent insight generation. Finally, “Proving AI ROI in the Enterprise” addresses the critical challenge of demonstrating measurable business impact from AI initiatives, outlining robust metrics that resonate with executives and boards. Attendees will gain practical strategies for architecting real-time pipelines, integrating AI into analytics responsibly, and measuring the tangible value of AI investments, leaving them equipped to accelerate data-driven decision-making across their organizations.

Agentic AI systems are moving rapidly from experimentation into enterprise production environments, where reliability, security, and governance matter as much as model capability. While large language models enable powerful reasoning and tool use, many early agent implementations struggle when exposed to real-world constraints such as inconsistent data, regulatory requirements, latency, and operational failure modes.

This talk explores how enterprises can design agentic AI systems that operate predictably and safely in production. Drawing from real deployment experience, it focuses on practical design patterns for planning loops, tool orchestration, and decision control, rather than prompt chaining or isolated demos. The session will examine how agents manage state, select actions, and recover from partial failures, as well as how guardrails—such as policy enforcement, observability, and human-in-the-loop controls—are implemented to reduce risk and maintain trust.

Attendees will gain a clear understanding of what differentiates experimental agents from enterprise-ready systems, along with concrete architectural considerations and lessons learned from deploying agentic AI in regulated, large-scale environments.

Abstract coming soon

From Code and Content to Insight and Action
Recent advances in AI have dramatically reduced the cost of generating code, analysis, and content. However, in financial systems and software engineering, the primary challenge is no longer production, it is quality. Faster and cheaper outputs do not necessarily translate into correct, reliable, or decision-relevant outcomes.

This talk focuses on how AI systems can be designed to improve the quality of outputs across the full pipeline, from code and data to analysis and decision-making. The objective is not to produce inexpensive approximations of human work, but to elevate standards: making outputs more accurate, more consistent, and more trustworthy.

We examine several technical limitations in current approaches and outline practical directions for improvement:

• Hierarchical and structural understanding
• Data quality, validation, and currency
• Semantic correctness in structured data
• Decision-aware reasoning and trade-offs

Across these examples, the central theme is that improving outcomes requires strengthening representation, validation, and reasoning—not simply increasing speed or scale. The next phase of AI is not about doing more, but about raising the standard of what is produced—enabling systems that support deeper insight, more reliable outputs, and higher-quality decisions

Financial institutions want AI/ML at scale, but brittle data pipelines, silos, and compliance demands slow progress. This talk shows how a Data Mesh—domain-oriented ownership, data-as-product, self-serve platforms, and federated governance—becomes the foundation for reliable, reusable ML features and trustworthy models. We’ll map mesh principles to FS use cases—fraud detection, risk, personalization—and show patterns for feature stores, lineage, quality, and access controls that satisfy regulators while accelerating delivery. Attendees will get a pragmatic blueprint: where to start, how to sequence capabilities, metrics that prove value, and pitfalls to avoid on the road from pilots to production.

Abstract coming soon.