Dialyn

01

The problem

Indian e-commerce loses money on unverified cash-on-delivery orders, failed payments and missed deliveries. Calling every customer by hand doesn't scale, and generic voice bots don't return data a store's backend can act on.

02

Key features

• 01Six order events (order confirmation, COD verification, payment success and failure, shipped, out for delivery), each with its own prompt, greeting and allowed outcomes
• 02Structured outcomes through a record_outcome tool. The agent states only the order facts it was sent and never promises changes
• 03Scheduler with per-merchant calling hours, time zones, retries on busy or no-answer, and concurrency limits
• 04Signed call.completed webhooks (HMAC-SHA256) with transcript, timeline and a stereo recording
• 05English and Hinglish (Sarvam Indian voices); per-merchant voice overrides, including consented ElevenLabs clones and recorded human greetings
• 06Shopify and WooCommerce webhooks trigger calls automatically, one call per order per event
• 07Twilio, Plivo and Exotel telephony, plus a browser WebRTC test call that needs no phone
• 08Multi-tenant merchant dashboard: API keys, team roles, analytics, call log and call detail

03

Architecture highlights

A FastAPI control plane (merchant API, scheduler, webhooks) dispatches calls into per-call Pipecat pipelines: VAD → STT → LLM with tools → TTS, streamed over telephony WebSockets or WebRTC.

• Cascade pipeline by default, so every stage is swappable per merchant through config, including speech-to-text, LLM and text-to-speech providers.
• Call state lives in the database, not the worker, which keeps voice workers stateless.
• Idempotency keys on call creation and one-call-per-order-per-event on store webhooks prevent duplicate calls.
• 51 pytest tests cover the API, merchant flows, platform integrations and WebRTC networking.

04

Tech stack