VoiceOps

Inspiration

Production incidents are stressful and expensive. Engineers often jump between dashboards, logs, GitHub, terminals, and deployment tools just to understand what broke. We wanted to explore a simpler workflow: what if developers could talk directly to production and let AI handle investigation and execution?

That idea became VoiceOps, an AI on-call engineer that turns incident response into a conversation.

What it does

VoiceOps allows engineers to interact with infrastructure and codebases through voice commands.

A developer can say:

“Latency increased after the latest deployment, investigate and prepare a fix.”

The system listens, understands intent, gathers logs and deployment context, analyzes the codebase, generates a patch, opens a pull request, and prepares a deployment while keeping the human in control of final approval.

How we built it

We designed VoiceOps as a multi-layer agent architecture:

Voice interface for natural language interaction
Agent orchestrator for planning and execution
MCP-based action layer to interact with developer tools
Repository analysis and code generation pipeline
Deployment and observability workflow

Core flow:

Voice → Agent → MCP Actions → GitHub → Deploy → Verification

The system retrieves runtime context, identifies likely root causes, proposes fixes, and executes recovery workflows.

Challenges we ran into

The biggest challenge was balancing automation with safety.

Allowing an agent to modify production-related code introduces trust and validation concerns. We solved this by introducing approval checkpoints before deployment.

Another challenge was connecting multiple tools into a reliable workflow while preserving context across debugging, code generation, and deployment steps.