Nirdesh AI
LiveNirdesh AI is an instrumented multilingual voice agent built as a sandwich architecture: streaming speech-to-text feeds a LangGraph-backed text agent, which pipelines into streaming text-to-speech returning 24 kHz PCM audio over an authenticated WebSocket.
Nirdesh AI features & system architecture:
- Multilingual voice loop using Saaras v3 STT, sarvam-105b LLM, and Bulbul v3 TTS
- Silero ONNX VAD running 32 ms local frames, speech pre-roll, 500 ms endpointing, and silence gating
- Barge-in handling that cancels generation, tears down active TTS sockets, flushes audio, and truncates history
- LangChain tools for server-side Exa web search and browser-proxied local todo management
- Paise-denominated wallet, mock recharge, per-second voice-session billing, warnings, and auto-cutoff
- Full-stack instrumentation with TurnTimer metrics, structured logs, and client latency waterfalls
A production-style voice platform offering real-time multilingual voice interaction with low-latency streaming and full-stack session instrumentation.
Stack used
FastAPILangGraphPythonReactTypeScriptDockerTailwind CSS