Skip to content
SQ

2026Enterprise conversational AI over the phone

Voice Agent "Mariam"

A realtime multilingual voice agent for inbound and outbound calls: lead qualification, FAQ handling, routing, appointment booking, transcription, summarisation and WhatsApp follow-up, with CRM-style workflows behind it.

Status · Built as an enterprise voice agent. The demo on this site is an architecture visualisation — no live call backend is connected here.

The problem

Businesses lose qualified inbound because nobody answers the phone, and waste senior time on calls that were never going to qualify. An IVR tree is not an answer: callers abandon it, and it cannot ask the one follow-up question that decides whether a lead is real.

The approach

  1. 01Build a streaming voice pipeline — speech in, LLM reasoning, speech out — tight enough that turn-taking feels like a conversation rather than a walkie-talkie.
  2. 02Give the agent the jobs a receptionist actually does: qualify, answer known questions, route, book, take a message.
  3. 03Ground answers in the business's own documents so the agent does not improvise policy.
  4. 04End every call with structured output — transcript, summary, qualification fields, follow-up — pushed into CRM-style workflows and WhatsApp follow-up.

What I did

  • Designed and built the realtime pipeline and the conversation state machine.
  • Integrated telephony, streaming speech recognition and speech synthesis into a single low-latency loop.
  • Built retrieval grounding over business documents and the structured post-call extraction.
  • Implemented multilingual handling and the escalation rules that hand a call to a human.

Architecture

Tap any component to see what it does and what it runs on.

Ingestion

Processing

Models

Orchestration

Storage

Interfaces

Select any component to inspect it. Architecture of the production system. Nothing on this website places or receives calls.

Trade-offs I would defend

Latency is the product

Every architectural choice — streaming at each hop, partial transcripts, early TTS start — exists to protect turn-taking. A smarter model that adds a second of silence makes the agent worse, not better.

Narrow scope, honest escalation

The agent is scoped to qualification, FAQ, routing and booking. Outside that it escalates. Letting a voice agent improvise on commercial terms is where these systems embarrass their owners.

Stack & evidence

  • Pipecat
  • Twilio
  • Deepgram
  • Cartesia
  • pgvector
  • Redis
  • Django
Role
Architecture and implementation
Stack
Pipecat, Twilio, Deepgram, Cartesia, pgvector, Redis, Django
Live demo here
Not connected — architecture visualisation only

Next case study

TCG Demand Forecasting

Hybrid ML + LLM forecasting system