Miguel Otero Pedrido

In this workshop, we’ll build a fully functional multimodal Telegram agent, putting into practice a wide range of concepts from the world of Agentic AI. This isn’t just another PoC — it's designed for those who are ready to level up and build complex, production-ready agentic applications. 

Throughout the session, you’ll learn how to build a Telegram agent you can chat with directly from your phone, master the creation and management of workflows with LangGraph, and set up a long-term memory system using Qdrant as a vector database. We’ll also leverage the fast LLMs served by Groq to power the agent’s responses, implement Speech-to-Text capabilities with Whisper, and integrate Text-to-Speech using ElevenLabs. Beyond language, you’ll learn to generate high-quality images using diffusion models, and process visual inputs with Vision-Language Models such as Llama 3.2 Vision. 

Finally, we’ll bring it all together by connecting the complete agentic application directly to Telegram, enabling a rich, multimodal user experience. Throughout the day, you will focus on the following key areas:

• Understand the full architecture and stack for building production-grade multimodal agents.
• Learn to build and debug agent workflows using LangGraph and LangGraph Studio.
• Implement short-term (SQLite) and long-term memory (Qdrant) systems for your agent.
• Enable speech interactions using Whisper (STT) and ElevenLabs (TTS).
• Integrate vision-language understanding with Llama 3.2 Vision and generate images via diffusion models.
• Connect your agent to Telegram for real-time, mobile-accessible interactions.

In this workshop, participants will work hands-on with a cutting-edge stack of tools and technologies tailored for building multimodal, production-ready agentic applications. LangGraph serves as the backbone for orchestrating agent workflows, with LangGraph Studio enabling easy debugging and visualization. SQLite powers short-term memory within the agent, while Qdrant, a high-performance vector database, handles long-term memory for contextual awareness. Fast and efficient responses are delivered using Groq LLMs, complemented by natural voice interactions through Whisper for speech-to-text and ElevenLabs for text-to-speech synthesis. For visual intelligence, Llama 3.2 Vision interprets image inputs, and diffusion models are used to generate high-quality visuals. Finally, the complete system is integrated with the Telegram Bot API, allowing users to interact with the agent in real time via chat, voice, or image directly from their mobile devices.

Prerequisites:

• Basic Python programming skills
• Familiarity with LangChain or LangGraph
• Basic understanding of multimodal AI concepts

*Note: These are tentative details and are subject to change.

In this hands-on session, we'll move beyond demos and PoCs to dive into how to build complex agentic systems that work in real-world scenarios. We’ll start by covering the fundamentals of agents (short-term memory, long-term memory, tool use, reasoning techniques, etc), then introduce Agentic RAG and how it differs from traditional RAG, and show how to bring these concepts into production using LLMOps practices like agent monitoring, prompt versioning, dataset management and RAG evaluation. We'll wrap up with a real-time simulation of agents operating inside a video game, seeing all these concepts come to life in action.