Hyderabad,

August 14

1.00pm - 1.40pm
Deconstruct, then reconstruct: AI-first legacy migration

Muskan Dhanda and Ashish Saini

Explore an agentic orchestration pipeline for large-scale legacy modernisation that shifts the focus from code translation to behavioural reconstruction. Extract language-agnostic specifications covering business rules, data models, architecture, and cross-cutting concerns to deterministically regenerate applications on a new technology stack. Stateful test-driven generation, automated validation, and iterative self-correction ensure behavioural equivalence while reducing migration effort and risk. The process also surfaces undocumented business logic and edge cases, giving teams deeper insight into legacy systems.

1.40pm - 2.20pm
Engineering the harness
Jaya Simha Reddy Nandyala and Prabina Pani

Your AI coding agent isn't the problem, the harness around it is. In "Engineering the Harness," we explore why smart models ignore team conventions, then share a practical system of guides that steer agents before they act and sensors that catch mistakes after. Drawing from a brownfield multi-repo platform and a greenfield agentic workflow, we show how to build a setup that works with any AI tool or stack. Leave with a concrete checklist to audit your team's AI setup.

2.20pm - 3.00pm
Fireside chat with Dr. Praveen Gummadidala

An innovator at heart, Dr. Praveen brings 22+ years of experience turning ambitious ideas into multi-million-dollar realities and building game-changing products across tech-heavy industries. A global technology leader and AI aficionado, he combines deep expertise in AI, ML, deep learning, computer vision, and GenAI with a Ph.D. in Applied AI. Recognized for innovation and excellence, he is also a passionate mentor who has guided 800+ professionals worldwide.

1.00pm - 1.40pm
But it worked in my PoC!
Divye Singh and Madhukar Bhandari.

What does it take to automate enterprise IT at scale with AI agents? More than expected. This talk shares the hard-won lessons from deploying agents to manage software updates across thousands of devices globally. Our first attempt collapsed under bloated prompts. The second relied on an all-knowing orchestrator that consumed 80% of its context, took two minutes to respond, burned 3M tokens per run, and hallucinated its way to failure. The third shipped.

Discover what finally worked: smaller agents with tighter responsibilities, on-demand skills, structured agent-to-agent communication, dynamic agent spawning, deterministic workflows, dedicated Q&A, and tool-call validation. Learn why multi-agent systems are a distributed systems problem first and an AI problem second. No theory. Pure scar tissue.

1.40pm - 2.20pm
Deep research at scale: Evolving from LangGraph to DeepAgents across 15M+ documents
Sindhu Rathlavath and Ishaan Bhatnagar

Scaling LLMs to 15M+ domain-heavy documents exposes the limits of standard RAG. Complex enterprise queries demand multi-hop reasoning, multi-document synthesis, and strict domain accuracy.

We share our blueprint for taking a multi-agent deep research engine from prototype to production, covering HPC-based retrieval, the evolution from LangGraph ReAct to DeepAgents, and evaluation using reranking, LLM-as-a-judge, custom rubrics, and human verification. We also share lessons from MLflow observability, streaming logs, and Bedrock infrastructure.

Learn what it takes to reliably bring non-deterministic agentic workflows into production at scale.

2.20pm - 3.00pm
TFM based Shared Intelligence Layer - Augmenting your existing AI/ML models
Muralikrishnan Puthanveedu, Shrijayan Rajendran and Mitra Bhanu Rath (NVIDIA)

Every model in a bank sits on its own feature pipeline, compressing the same customer event stream into aggregates that discard ordering and timing. Transaction Foundation Models learn from the sequence instead. This session explores what a shared intelligence layer changes for models already in production, what augmenting existing models involves, and how the NVIDIA stack and blueprint make the build methodical and repeatable. Includes a live demonstration.

1.00pm - 1.40pm
Human in the loop is not enough — real lessons from securing an autonomous AI platform
Uday Kumar S S, Navin Govindarasu and Piyush S

Shipping an AI product is easy; securing it is where teams get caught off guard. We share lessons from building an enterprise AI platform on a third-party foundation model, where human oversight proved insufficient as the platform evolved toward autonomous and agentic capabilities.

Explore practical security lessons across governance, risk, cloud architecture, threat modelling, penetration testing, and guardrails for systems that act, not just respond. See how risks evolved from OWASP Top 10 for web apps to LLM and Agentic Apps.

No theory. Just real decisions, gaps, and what we would do differently. Leave with a practical security framework for AI platforms and answers to the hard questions before someone else asks them.

1.40pm - 2.20pm
One guardrail is not enough: Defense-in-depth against LLM jailbreaks
Nitin Gupta

Production chatbots have been talked into binding contracts, creating legal liability, and leaking data through RAG. OWASP ranks prompt injection as the #1 GenAI risk and notes that RAG and fine-tuning do not eliminate it.

This talk presents a defense-in-depth architecture on AWS, reframing jailbreaks as two problems: a universal safety baseline inherited from the model and a bot-specific scope envelope that teams must define. Explore the pipeline: a fine-tuned BERT classifier on SageMaker, moderation and PII filtering, an LLM canonicalizer, hostile-input retrieval, guarded generation, and output filtering. Close the loop with asynchronous LLM-as-judge triage that turns production attacks into training data, continuously strengthening the cheapest defense layer.

2.20pm - 3.00pm
The AI Kill Chain: How small AI weaknesses become enterprise breaches
Nalinikanth Meesala

AI vulnerabilities rarely cause major incidents alone. Impact emerges when weaknesses chain across AI systems, identities, tools, data, and business processes.

This talk introduces the AI Kill Chain through realistic scenarios and a live demo, showing how prompt injection enables tool abuse, poisoned knowledge influences decisions, agent memory creates persistence, and agent trust enables lateral movement. See how least-privilege agents, tool authorization, trust boundaries, human approval, and secure architecture can break the chain.

Leave with a practical methodology to identify, model, and mitigate AI attack chains before they become breaches.