Skip to content
Start a project
Start a project

[email protected]LinkedIn

Client projectClient confidential

A hands-free assistant that listens and replies in natural Hindi, in real time.

Services
AI Engineering
Client
Confidential
Built with
Python · Pipecat · Deepgram · Llama 3.3 · Groq · OpenRouter · Sarvam AI · FastAPI
The assistant's landing page: an AI voice assistant that speaks your language, with natural Hindi speech and real-time replies.

01

The client needed a voice assistant that understands and speaks natural Hindi, fast enough to feel like a real conversation.

02

  1. 01A streaming pipeline from speech to text, through a large language model, and back to speech.
  2. 02Deepgram Nova-3 for transcription and Sarvam AI Bulbul v3 for natural Hindi speech.
  3. 03Llama 3.3 70B, served through Groq and OpenRouter, for conversation and reasoning.
  4. 04Streaming transcripts and word-by-word replies, so answers start sooner and conversations feel natural.
  5. 05An asynchronous FastAPI backend that coordinates listening, thinking and speaking.
Built with
  • Python 3.13
  • Pipecat
  • PyAudio
  • Silero VAD
  • Deepgram Nova-3
  • Llama 3.3 70B
  • OpenRouter
  • Groq
  • Sarvam AI Bulbul v3
  • FastAPI
  • Uvicorn
  • WebSocket
  • python-dotenv
  • Loguru
  • HTML
  • CSS
  • JavaScript
  • Canvas
  • Google Fonts

The product

A voice conversation in progress: the assistant listening, and its spoken reply with a three-day travel plan for Jaipur.
Chat history: past voice conversations, each with a title, a short summary and a play button.
The playground: the speech-to-text, language and text-to-speech models (Deepgram Nova-3, Llama 3.3 70B on Groq, Sarvam AI Bulbul v3) and example prompts.

Representative interface; the client's product is confidential.

Next project

Client confidential

Voice Call & Document Summarisation

Read the case study