Mumbai Artifical Intelligence Group

Vision-Language Models (VLMs) are transforming how AI sees, understands, and explains the world. From image captioning to multimodal reasoning, they power the next wave of intelligent applications.
In this live session, we’ll:

• Break down the fundamentals of how Vision-Language Models process both text and images
• Showcase lightweight VLMs (SmolVLM, MiniCPM-o, Qwen-VL, etc.) that can run on modest hardware
• Demonstrate real-time examples: image captioning, visual Q&A, and multimodal retrieval
• Compare open-source VLMs with larger commercial ones, and discuss where small models shine
• Share tips on deploying VLMs for startups, apps, and research projects

🔹 Format: Demo-driven walkthrough + interactive Q&A
🔹 Who’s it for: AI engineers, product managers, researchers, and builders curious about multimodal AI
🔹 Takeaway: A working understanding of VLMs, access to demo notebooks, and ideas for real-world applications

Retrieval-Augmented Generation (RAG) doesn’t stop at text. The future is multimodal RAG, where models can reason over documents, images, charts, and more.
In this hands-on session, we’ll explore:

• What Multimodal RAG is and why it matters
• How to combine text + images in a retrieval pipeline
• Using vision-language embeddings for storing & searching multimodal data
• Running live demos with small VLMs (Vision-Language Models) and vector databases
• Practical use cases: compliance checks, document Q&A, product search, and research workflows

🔹 Format: Interactive demo + live coding walkthrough
🔹 Who’s it for: AI engineers, researchers, and product teams building advanced AI systems
🔹 Takeaway: A working notebook + examples of multimodal retrieval powering next-gen AI apps

Artificial intelligence agents are rapidly evolving, but how do we make them learn better behaviors in complex real-world environments? In this session, we’ll explore ART (Agent Reinforcement Training) from OpenPipe, a framework that uses reinforcement learning (RL) to refine how AI agents make decisions, handle uncertainty, and improve over time.
We’ll break down:

• What ART is and why it matters for agent performance
• How reinforcement learning techniques can improve reasoning and adaptability
• Practical examples of training agents with ART using OpenPipe
• Real-world use cases: from customer support agents to autonomous workflows

This event is ideal for engineers, researchers, and startup builders who want to push beyond static prompt engineering and learn how to apply reinforcement signals to build smarter, more reliable AI agents.
Join us for an interactive demo and discussion on how ART can transform your AI development workflow.

AWS GenAI Builder Series — AWS AgentCore Services
Format: 2 hours (presentation → live demo)
Focus: Build and operate AI agents on AWS using Amazon Bedrock AgentCore. We’ll work directly from the official AgentCore Samples repo to take an agent from local run to a production-ready deployment pattern.
What you’ll learn

• Where AgentCore fits in an AWS stack: Runtime, Gateway (tooling), Memory, Identity, and Observability
• How to run an agent locally, then deploy it on AWS
• How to register and call tools (Lambda/HTTP) through the Gateway
• How to add state with Memory and secure access with Identity
• Practical ops: tracing, debugging, guardrails, and cost awareness

Live demo (using the GitHub samples)

1. Runtime quickstart: run a minimal agent locally → deploy → invoke
2. Gateway tool call: expose a simple Lambda/HTTP tool and call it from the agent
3. Add Memory + Observability: persist context and trace steps end-to-end
4. (If time) wire up Identity for a third-party call

Who should attend
Engineers and architects building agentic applications on AWS (SWE, MLE, SRE, platform).
Prerequisites (bring your laptop)

• AWS account with access to Bedrock and AgentCore
• AWS CLI configured (aws configure)
• Python 3.10+ and Docker/Finch for local dev
• Access to at least one Bedrock model (e.g., Claude)
• We’ll reference: awslabs/amazon-bedrock-agentcore-samples

Agenda (2 hours)

• 15 min — Context & goals
• 35 min — Theory walkthrough (architecture, patterns, gotchas)
• 50 min — Live demo (zero → working agent)
• 10 min — Q&A + pitfalls
• 10 min — Homework lab & repo pointers

Takeaways

• Slides and a reproducible demo repo
• A mini runbook for local → cloud deployment
• A checklist to productionize your first AgentCore agent