Sai Byrraju
< Back to projects

VoiceOps

Voice-driven production incident response.

PythonGoogle ADKGeminiDynatrace MCPFastAPIVAPITwilioGitHub ActionsGoogle Cloud Run

Why I built this

For the Google Rapid Agent Hackathon, I built an agent that detects a production problem, calls the on-call engineer, explains it, gets verbal rollback approval, and executes it.

How it works

dynatrace alert → agent → github correlation → vapi voice call → approval → github actions rollback → post-incident analysis

Dynatrace alerts are correlated with recent GitHub commits. VAPI calls the on-call engineer; approval triggers GitHub Actions to build and deploy the rollback SHA. After mitigation, DQL analysis, anomaly detection, blast-radius calculation, and a report run. A browser dashboard can resolve the same approval.

Decisions

Voice call as primary interface

A phone reaches an engineer away from a laptop; the dashboard is the fallback.

Two-phase workflow

Mitigate first, analyze second. The agent should not run a full RCA while production is still broken.

Resolve parent SHA through GitHub

trigger_github_rollback finds the bad commit parent through the API instead of hardcoding a revision.