Build Advanced AI Agents: Voice, Research & Multi-Agent Workflows

Building Advanced AI Agents

• 💡 This workshop series focuses on building three types of advanced AI agents: voice agents, deep research tools, and multi-agent workflows.
• 🚀 Participants will gain hands-on experience with popular AI tools and frameworks, including LiveKit, Exa, LangChain, and Cerebras.
• 🛠️ The series provides sample code, practical exercises, and open-source repositories to facilitate building functional AI agents.

Workshop 1: Voice Sales Agent with LiveKit

• 🗣️ Learn to build a real-time voice sales agent capable of natural conversations, pulling product context, and responding intelligently.
• 🧠 The agent's pipeline involves Automatic Speech Recognition (ASR), Voice Activity Detection (VAD), Large Language Models (LLMs) for thinking, and Text-to-Speech (TTS) for speaking.
• ⚡ Cerebras provides ultra-fast inference, reducing response times from seconds to milliseconds, crucial for natural conversation.
• 🤝 LiveKit acts as middleware, managing audio streams, context, and orchestrating AI services.
• 💬 Cartisia's Ink engine handles speech-to-text and text-to-speech with real-time accuracy and low latency.
• 🧩 Retrieval Augmented Generation (RAG) is used to load business-specific information, like product descriptions and pricing, into the LLM's context to minimize hallucinations.
• 👥 The system can be expanded into a multi-agent system with specialized agents for technical support or pricing inquiries, enabling seamless handoffs.

Workshop 2: Deep Research Assistant with Exa

• 🔍 Build an AI research assistant that can autonomously search the web, analyze multiple sources, and provide structured insights in under 30 seconds.
• 🌐 Exa's search API is used to find relevant web content, returning full page content for comprehensive analysis.
• 🧠 The process involves breaking down research questions, using RAG to provide LLMs with context, and identifying knowledge gaps for follow-up searches.
• 🚀 Cerebras' fast inference is critical for chaining multiple LLM calls, enabling complex research tasks to be completed rapidly.
• 🧩 A recursive approach is implemented where the LLM identifies gaps and triggers new searches, leading to richer, more comprehensive results.
• 🧑‍🔬 Advanced implementations can involve a lead agent breaking down queries into subtasks for multiple specialized agents working in parallel.

Workshop 3: User Research Automation with LangChain

• 📊 Automate user research by generating user personas, conducting simulated interviews, and synthesizing feedback in under 60 seconds.
• 🤖 The system uses AI to generate interview questions, create diverse AI personas, run simulated interviews, and analyze responses.
• ⏱️ This process compresses weeks of traditional user research into minutes, significantly accelerating innovation speed.
• ⚙️ LangGraph is used for workflow orchestration, defining nodes (Python functions) for each research step, and managing state across the process.
• 🧠 Cerebras powers the LLM for fast generation of personas, interview questions, responses, and insights.
• 🤝 Specialized agents (nodes) handle tasks like configuration, persona creation, interviews, and synthesis, updating a shared state object for coherent workflows.
• 📈 The system can be expanded with multi-question interviews, follow-up questions, conversation memory, and enhanced synthesis for deeper insights.