Meeting-to-PR: Voice-Driven Code Generation

Role: System Architect & Technical Lead

Overview

Meeting-to-PR is an AI-powered Discord bot that transforms verbal developer conversations into production-ready GitHub pull requests. Developers describe coding tasks in natural language; the system autonomously generates code changes, creates PRs, runs automated reviews, and applies fixes — streamlining the entire development workflow.

System Architecture

Built as a modular three-component pipeline:

Discord Voice Layer — Transcribes conversations using ElevenLabs and extracts task requirements via conversational AI
Coding Agent — Generates code changes in isolated Daytona sandboxes and creates GitHub PRs
Orchestration & Review — Coordinates the pipeline, manages automated reviews, and handles iterative fixes

My Contributions

Designed the complete system architecture and data contracts enabling parallel development across three team members
Architected state machine for task lifecycle: PENDING → CODING → CREATING_PR → REVIEWING → FIXING → DONE
Designed conversation flow and state machine for natural language task extraction
Designed central orchestration layer with auto-fix loop (max 3 attempts) and Sentry integration

Key Impact

Reduced conceptual code-to-PR workflow from 30+ minutes to under 2 minutes
Enabled true hands-free development through voice-driven task creation
Created clean component interfaces enabling parallel team development

Why I Built It

Writing code is rarely the bottleneck — context switching is. The idea was to remove the overhead between a verbal decision in a meeting and an actual PR, and use it as a real test of how far multi-agent orchestration could go with minimal human-in-the-loop steps.

Technologies

Python · asyncio · Discord.py · ElevenLabs · Daytona SDK · Claude API · PyGithub · Sentry