SciFi Agents
Multi-tenant Bangla AI voice agents over the phone
The question
Can a telephone call become a job another system posts over an API, answered by a Bangla agent that states only what it was given and falls silent the moment a human joins?
A platform where third-party systems post a job and a Bangla-speaking AI agent places or answers the call over SIP, runs a no-code-configured script, and reports a structured result to a signed webhook. Two reference agents: order confirmation for e-commerce and 999 emergency triage.
- 1
- Media hops
- 2
- Reference agents
a call as a job, answered by an agent that only states what it was given
What it is
Every agent is fully configurable without code: one authored prompt, a knowledge base it searches, the capabilities it is given, and the APIs it may call, all defined in an admin console. Every agent shares three hard guarantees: it only states facts it was actually given, it never acts on the third party's behalf, and once a human joins it goes quiet and answers only when addressed.
How it works
- Asterisk is the entire media plane: it terminates provider trunks, mixes each call in a bridge, carries audio to the worker over AudioSocket, and serves the console's web phone over WebRTC. One hop, one codec, no resampling on the agent path.
- The agent worker uses livekit-agents purely as a voice pipeline library (VAD, endpointing, barge-in, streaming STT, LLM, and TTS) with no LiveKit server. Per-party snoop channels keep speech attributable once more than one human is on the line.
- A scheduler on a Postgres-backed queue owns the job and call lifecycle: dispatch, retries, finalisation, recording upload, webhook delivery with HMAC signing.
- A grounding gate in the shared core enforces the guarantees above.
Status
First live end-to-end calls completed; the 305-second STT cap and long-call dropouts fixed; console, API, and SIP runbook documented.
Connections
Other work that sits in some of the same fields. The overlap is what the map draws.