Warta AI – Multimodal Conversational Companion
Multimodal conversational companion platform featuring low-latency real-time voice streaming, multilingual speech synthesis, emotional persona adaptation, and persistent Pinecone vector memory.
Real-Time Full-Duplex Voice Streaming
Architected a real-time bidirectional audio pipeline utilizing WebSockets and WebRTC audio channels. Coupled with low-latency Voice Activity Detection (VAD) and fast speech-to-text models, the platform achieves a sub-80ms Time-To-First-Token (TTFT), creating fluid, natural voice conversations without unnatural pauses.
Multilingual Companion Personas & Sentiment Tuning
Built dynamic persona modulation across multiple categories (Empathetic, Professional, Mature, Relationship, Storytelling) with native Hindi and English language support. The system dynamically tailors speech cadence, emotional warmth, and vocabulary based on live sentiment analysis.
Persistent Vector Memory & User Context Retention
User memories, conversational preferences, and historical dialogue threads are tokenized into Pinecone vector namespaces. Dynamic semantic retrieval injects past conversational context into active sessions, eliminating context drift across multi-day interactions.
Interactive Holographic Voice UI & Scalable Backend
Designed a responsive, tactile UI featuring a real-time reactive voice orb visualizer that animates synchronously with speech waveforms. Backend services are decoupled into Node.js microservices with Redis caching and MongoDB user profile storage.`
