Shubham Agnihotri
Chief Manager - Generative AI - IDFC Bank
Mumbai, India
Actions
Shubham Agnihotri is a Chief Manager – Generative AI at IDFC FIRST Bank, with 6+ years of experience working across diverse modalities like text, images, audio, video, tabular, and graph data. He specializes in fine-tuning and deploying Large Language Models (LLMs) and multimodal AI systems in production environments, contributing to real-world impact at scale.
Shubham is an active community leader and currently serves as an organizer for the AWS User Group Mumbai, where he helps nurture the cloud and AI ecosystem through meetups, hackathons, and hands-on sessions. Previously, he co-organized TensorFlow User Group (TFUG) Bangalore, leading over a dozen developer events.
He is also the founder of an agriculture drone startup that enabled precision spraying services for farmers. His solution helped save over 1 million liters of drinking water, reduce pesticide usage by 40%, and was recognized as an Emerging Agri-Drone Startup by the Times of India Group. This mission-driven venture drives his continued passion for building tech with societal impact.
🎤 Featured Speaking Engagements
AsiaTech Singapore -
1. https://asiatechxsg.com/techxlr8asia/speakers/shubham-agnihotri/
2. https://asiatechxsg.com/developerxperience-summit/speakers/shubham-agnihotri/
Aws Community Days Bangalore: https://acd.awsugblr.in/
Google Cloud Community Day: https://www.linkedin.com/feed/update/urn:li:activity:7337524380920348674/
TechShow London : https://youtu.be/jx_9SbARV9M?si=6PN1h1TYhgAPeQ4z
AWS Community Day Mumbai : https://youtu.be/XihAhZQZtV4?si=ST8XYvuRTdgoArLq
In addition to community and enterprise work, Shubham is part of the Google AI Community Network and was the founding lead of Cynergy, the official coding group of Ramaiah Institute, where he fostered a peer-to-peer learning culture.
Links
Area of Expertise
Topics
Mastering Google's A2A Protocol: Deep Dive on Connecting AI Agents to Agents
In today's rapidly evolving AI landscape, the ability for autonomous agents to communicate and collaborate seamlessly is paramount. Google's Agent-to-Agent (A2A) protocol addresses this need by providing an open standard that enables diverse AI agents to interact securely and efficiently across various platforms and vendors. This session offers a comprehensive exploration of the A2A protocol, delving into its architecture, key features, and practical applications. Attendees will gain insights into implementing A2A in real-world scenarios, understanding its role in enhancing interoperability, and leveraging it to build robust, multi-agent ecosystems.
Key Takeaways:
Understanding A2A's Architecture:
Gain a deep understanding of the A2A protocol's design, including its communication mechanisms, security features, and scalability considerations.
Implementing A2A in Real-World Scenarios:
Learn how to integrate A2A into existing AI systems, facilitating seamless agent-to-agent communication across diverse platforms.
Enhancing Interoperability:
Explore how A2A promotes interoperability among AI agents, enabling them to work collaboratively regardless of their underlying technologies.
Comparative Analysis with MCP:
Understand the differences and complementarities between A2A and Anthropic's Model Context Protocol (MCP), and how they can be leveraged together for more robust AI ecosystems.
Building Multi-Agent Ecosystems:
Discover best practices for constructing scalable and secure multi-agent systems using A2A, and how to address common challenges in agent communication.
Autonomous Knowledge Agents with Google MCP toolbox, Gemini, Vertex AI, and Neo4j
In the rapidly evolving landscape of artificial intelligence, the integration of large language models (LLMs), agent orchestration frameworks, and knowledge graphs is transforming how systems process and reason over complex data. This session delves into constructing autonomous knowledge agents by integrating Google's Gemini LLMs, Vertex AI's Agent Builder, Neo4j's graph database, and the Model Context Protocol (MCP) via the MCP Toolbox.
Attendees will explore how to harness Gemini's advanced language understanding, orchestrate multi-agent workflows with Vertex AI, and leverage Neo4j's graph structures for contextual data representation. The talk will also cover the implementation of GraphRAG (Graph Retrieval-Augmented Generation) techniques, enabling agents to retrieve and reason over structured knowledge effectively.
Through real-world examples and demonstrations, participants will gain insights into building scalable, explainable, and efficient AI agents capable of autonomous decision-making and continuous learning.
Key Takeaways:
1. Leveraging Google Gemini for Advanced Language Understanding:
Understand the capabilities of Gemini LLMs in interpreting complex queries and generating context-aware responses.
Learn how to integrate Gemini with agent frameworks to enhance natural language interactions.
2. Orchestrating Multi-Agent Workflows with Vertex AI Agent Builder:
Discover how Vertex AI facilitates the development and deployment of AI agents.
Explore tools like the Agent Engine and MCP Toolbox for managing agent interactions and workflows.
3. Utilizing Neo4j for Contextual Knowledge Representation:
Learn how to model and store complex relationships using Neo4j's graph database.
Understand the role of knowledge graphs in providing context and enhancing agent reasoning.
4. Implementing GraphRAG for Enhanced Information Retrieval:
Explore the GraphRAG approach to combine retrieval-augmented generation with graph databases.
See how agents can retrieve relevant information from Neo4j to inform their responses.
Case study - Unlocking Customer Insights: AI-Powered Conversation Intelligence for Banks
Customer interactions hold a wealth of insights, but manual evaluation is slow, inconsistent, and reactive. This session reveals how AI-powered conversation intelligence transforms raw call audio into actionable business insights. By leveraging Google’s advanced speech-to-text and emotion-analysis technologies, Neo4j graph databases, and Generative AI, we automate the identification of customer pain points, sentiment trends, and agent performance metrics—enabling real-time, scalable quality assurance and continuous coaching.
We’ll walk through a cross-functional pipeline: from live call ingestion and transcription, sentiment and intent analysis, to structured graph storage in Neo4j, and finally visualization of interaction patterns and agent scorecards. Attendees will see how this framework equips managers with real-time dashboards, automated evaluation frameworks, and triggers for root-cause investigations—raising the bar for customer care excellence.
Key Takeaways
- From Voice to Insights via Google’s Audio Tech
# Learn how to use Google’s speech-to-text API, audio emotion detection, and sentiment analysis to extract intent, emotion, and language patterns from live customer calls.
# Understand the importance of accurately capturing both what customers say and how they say it.
- Modeling Conversational Analytics in Neo4j
# See how to structure call metadata, speakers, intents, sentiment flows, and tags within a Neo4j graph
# Discover why relationships (e.g., call → query type → agent response) empower pattern discovery and timely insight alerts .
- Automated Detection of Pain Points & Coaching Needs
# Learn how AI identifies recurring customer complaints, root-causes, and potential service failures automatically
# Dive into agent performance scoring frameworks—measuring factors like empathy, response accuracy, and compliance—at scale.
- Live Quality Scoring & Manager Dashboards
# Understand how real-time dashboards ingest graph updates from Neo4j to reflect emerging trends, agent scoring, and customer sentiment distribution.
# Explore alerting systems that notify managers of low-quality exchanges or rising negative sentiment clusters.
- Scalable, Explainable, and Continuous Improvement
# See how graphs enable explainability: managers can trace a low score to specific call segments and transcripts.
# Learn how the pipeline supports continuous learning—enhancing speech models, conversation flows, and agent training based on data-driven insights.
Autonomous Knowledge Agents: Building a Next-Gen RAG System with AWS Bedrock, Claude, Lambda & S3
Step into the future of AI-powered retrieval with Agentic RAG—where LLMs don’t just fetch data, but think, reason, and act autonomously. In this session, we’ll harness the power of AWS Bedrock, Claude, Lambda, and S3 to build a next-gen self-improving retrieval system.
Key takeaways:
- Claude on Bedrock for dynamic, context-aware generation.
- S3 as a scalable knowledge vault for seamless retrieval.
- Lambda-driven orchestration for real-time, event-driven processing.
- Agentic workflows that refine retrieval and response autonomously.
Accelerating Training with Multi-GPU: Using PyTorch Lightning for PaLI-Gemma VLM
Dive into the world of high-performance model training with multi-GPU setups using PyTorch Lightning, featuring the PaLI-Gemma Vision-Language Model (VLM). This session will cover the essential tools and techniques for setting up efficient distributed training pipelines, allowing you to fully utilize multiple GPUs to speed up and scale model training. We’ll explore PyTorch Lightning’s seamless integration for handling multi-GPU workloads, making it accessible even for those new to distributed training. Attendees will get a step-by-step guide to configuring PaLI-Gemma for accelerated training, from data handling to model synchronization and optimization strategies. By the end, you'll have the skills needed to harness the power of multi-GPU training to scale multimodal AI models effectively.
MultiModel RAGs: Unlocking the Power of Multimodality with PaliGemma
Unlocking the Potential of Multimodality with Pali-Gemma. In this session, I’ll dive into the concepts and applications of multimodality, showcasing how combining multiple forms of data can enhance AI model capabilities. I’ll walk through the process of training on Colab/Kaggle, highlighting each step from data preparation to fine-tuning. Following the training phase, I’ll demonstrate how we can leverage Hugging Face Spaces for inference and deployment, providing a practical approach to deploying powerful Vision-Language Models (VLMs) in real-world applications. This session will serve as a comprehensive guide for anyone interested in building and deploying multimodal AI solutions, showing how PaLI-Gemma, a state-of-the-art model, is trained and implemented. By the end, participants will have a solid understanding of how multimodal RAGs work and how to bring them to production efficiently.
Building Transformers: Code Your Vision
From hugging face transformers to transformers from scratch in python, hands on Workshop with Transformers brush ups to implementing it.
Global Data & AI Virtual Tech Conference 2025 Sessionize Event
Cloud Community Days 2025 Mumbai Sessionize Event
AWS Community Day Bengaluru 2025 Sessionize Event
Devfest Mumbai, 2023 Sessionize Event
Shubham Agnihotri
Chief Manager - Generative AI - IDFC Bank
Mumbai, India
Links
Actions
Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.
Jump to top