Roshni is an ultra-low latency, real-time multilingual financial voice assistant for Bharat designed to assist users in English, Hindi (Devanagari), and Hinglish with a focus on financial inclusion in India.

Roshni delivers natural, human-like conversations with ultra-fast TTFB around 95 ms and latency under one second, while enforcing Devanagari script to prevent mispronunciations from romanized Hindi.

The system runs on a four-pipeline architecture that achieves sub-second latency, routing user WebRTC audio through LiveKit to concurrent processing by Deepgram Nova-3 for speech-to-text, Google Gemini as the language model, and Murf Falcon for text-to-speech, all coordinated with an in-memory SQLite store for analytics and persistence.

A two-agent design assigns general banking tasks to Roshni and government schemes (PM Mudra loans, subsidies, welfare programs) to Vikram, enabling domain-specific expertise and seamless handoffs with real-time transfer of conversation history during escalations.

Key capabilities include ultra-realistic Indian voices via Murf Falcon, native Devanagari script enforcement to preserve pronunciation, persistent memory of caller identity and preferences, tool execution for live rate lookups (e.g., fixed deposits and government schemes), and a structured multi-agent handoff workflow for routing users to the right specialist.

The project focuses on financial inclusion by removing language barriers, simplifying access to information on schemes and rates, and delivering a responsive, conversational experience tailored for regional and tier-2/3 markets in India.