Nov 21, 2026

Vigil: the AI on-call engineer that calls you with the fix, not the alert

agora conversational ai voice ai murf ai ai agents mcp incident response devops sre next.js

When production breaks at 3 AM, today's tools send an alert that only says something is wrong. Vigil investigates first, then phones the on-call engineer with the root cause and a safe fix, on a real-time voice call.

Why it matters: a customer-facing incident costs about $794,000 on average (PagerDuty, 2024), and most of that time goes into finding the cause. Vigil does that part before it wakes anyone, so the engineer starts with the fix.

What it does on one incident:
• Investigates before calling: alert, metrics, deploys, logs and runbooks. It traces the failure down the dependency chain, rules out other suspects, checks whether it hits one region or everyone and what customers see, runs a five-whys root cause analysis, and checks every possible fix for safety.
• Calls only when your own threshold is crossed: each team sets its alert level per metric.
• On the call: explains the cause down to the release, PR, commit and stack line, and answers "what changed?", "what did you rule out?", "who is affected?" and "what if?". Interrupt it any time, or type.
• Acts like a careful senior SRE: it changes production only after the engineer's spoken yes, which the server checks. Every change gets a change record and goes to 10% of traffic first; Vigil watches the canary, rolls out or reverts, and says the result out loud. It never changes code.
• Proves recovery from live metrics and says so unprompted, then closes out in order: status page, resolve, a postmortem with the root cause analysis and customer and region impact, action items filed as real GitHub issues, and a downloadable call transcript.

Also: English, Hindi and Hinglish in the same call. Teammates join from their phone, listen, and ask by voice or text, while only the on-call engineer approves changes. Vigil remembers past incidents, stays on topic, and reconnects its own voice if it ever goes silent.

Built on Agora's Conversational AI Engine: ARES speech recognition (English and Hindi together), GPT-4.1 mini, 13 tools over MCP or HTTPS, and the think and speak APIs. Every word Vigil says is Murf Falcon 2.1. Thali, the company in the demo, is simulated, so anyone can break production safely.

Try it live, no sign-up: pick an incident, see what's at stake, then break production. Or watch the demo of a real call.

Built solo by Yash Soni for Build with Agora, Bengaluru 2026 (PS31: From Workplace Assistance to Intelligent Action).

This build was uploaded as a hackathon project

Hackathon

Build with Agora - BLR

View All Projects
Give a star to encourage!Discussion
Start a new conversation!
Login to join the discussion