Be the on-call engineer
- Enter your Indian mobile number and the access code from the submission notes.
- We break one customer at random (you won't know which). Your phone rings in ~40 seconds from a US number.
- Pick up. Ask it anything: "what's going on?", "is it only one carrier?", "whose fault is it?"
- When it finds the cause it proposes a fix. Say "yes" and watch it apply and verify the fix below.
Calls are capped at 3 minutes; one call per number; numbers are stored only as a salted hash. Can't take a call? Watch without a call runs the same incident with a simulated phone.
Accounts (ranked by spend)
Call
Investigator, live (tool calls by the LLM; conclusions must pass code checks)
Raw CDRs (Plivo Zentrunk fields and hangup codes)
| time | from | carrier → network | code | cause | audio |
|---|
How it works
Per-account watch on failure rate, silent calls (answered but no audio) and volume drops. Most outages that matter don't show up as failures at all.
The moment it's suspected, an LLM investigator (Azure OpenAI) runs a triage pack and tool calls over CDRs, edge logs, CPS profiles and the account's change log. Ring time is free compute.
High-spend accounts get a phone call via Agora Conversational AI dialling out through a Plivo Zentrunk SIP trunk, voiced by Murf Falcon. Low-spend accounts get Telegram.
"Found it" is only allowed after code re-checks the conclusion on raw data. Wrong guesses get rejected and the agent keeps digging. It proposes the fix and the cost, then waits for your "yes".
Applies the fix (HA backup number, CPS increase, reroute, ACL entry, caller ID…), then watches traffic and says the before → after on the call.
Ticket with the investigation trail (with one deliberate failure + retry), customer email draft, account memory, summary.
Stack
- Voice: Agora Conversational AI (outbound telephony API, custom LLM endpoint, /speak break-ins) · Murf Falcon TTS
- Phone path: Agora → Plivo Zentrunk US trunk → Indian mobile (a US caller ID avoids India's domestic-anchoring 4590 rejection)
- Voice brain: our OpenAI-compatible proxy narrates the live case file each turn (Groq gpt-oss-120b, ~0.4 s to first word; Azure gpt-4.1-mini hedged backup)
- Investigator: Azure OpenAI gpt-5-mini tool-calling loop → gpt-4.1-mini → Groq → rules-only fallback; every conclusion verified in code
- Data: simulated fleet with real Plivo/Zentrunk field names and hangup codes (5350, 5180, 4170, 4590…). Fictional customers only.