The Most Overengineered Pocketwatch Ever Made

This session is a retrospective on the engineering behind a pocket watch that runs a real-time voice agent. Built end to end in a single week with Claude Code, from product idea to working reality complete with its own product page, this project showcases an entirely agentic workflow for rapid prototyping and iterative development. We'll explore the ESP32-S3 audio pipeline, including capture, echo cancellation, and the barge-in system that lets users interrupt mid-sentence, then follow the audio into Deepgram's Voice Agent API, where speech-to-text, LLM reasoning, and text-to-speech run as a single unified agent. The device becomes actionable when we target it at Azure OpenAI, or at your own ASP.NET Core endpoint hosted in Azure, so the agent's reasoning, function calling, and business logic live in the .NET stack you already know while the device stays lightweight.

Speakers

Ed Charbeneau