EVERY CHANNEL · FEATURED RAKE
Unisona
One AI agent, every channel.
What is it?
Build a multilingual agent once from your docs, website and spreadsheets, then deploy it to browser voice, phone, WhatsApp, Telegram, a website widget, a hosted chat page, email and the API, with one customer memory spanning every channel.
Why I built it
Businesses end up rebuilding the same assistant for every channel, and each copy forgets the customer. Unisona builds it once and remembers the customer everywhere.
What I actually built
- One agent deployed to browser voice, phone, WhatsApp, Telegram, a website widget, a hosted chat page, email and the API.
- One customer memory spanning every channel.
- Built-in CRM, campaigns, automations, human handoff, analytics, evaluations and a glass-box trace of every answer.
What's interesting about it
- Voice in 25+ languages with per-turn language detection (including Hinglish).
- Evals covering Ragas-style QA, a 12-attack red team, simulated customers, and voice simulations scoring WER and p95 latency. CI fails below 70%.
Architecture
Next.js web ↔ FastAPI (brain turn pipeline: guard → memory → RAG → LLM/tools → render; Pipecat voice: VAD + SmartTurn → STT → LLM → TTS), Supabase Postgres, Chroma + BM25.
Earlier version
Rebuilt and expanded from VoiceFlow AI: a multi-tenant voice SaaS with per-tenant RAG on ChromaDB + BM25, real-time Twilio/WebSocket streaming, FastAPI, LangGraph, Redis, PostgreSQL and MinIO.
- Next.js
- FastAPI
- Pipecat
- Supabase
- Chroma
- BM25
- LangGraph