Samvaad
Voice-first learning/information-gathering platform. Allows users to upload their docs and have text or voice conversations with them.
Problem Statement
Users struggle to quickly extract insights from large documents without reading everything. Traditional search is slow and voice-based interaction is limited.
Solution
Built a conversational AI that ingests documents and allows natural language queries via text or voice, using RAG for accurate responses.
Architecture
Frontend in Next.js handles UI and voice input/output. Backend uses FastAPI with OpenAI for embeddings and generation, Postgres for vector storage.
Problems Faced
Integrating ASR/TTS reliably, optimizing vector search for large docs, handling privacy concerns with user data.
What I Learned
Deepened knowledge in LLM fine-tuning, real-time audio processing, and scalable data pipelines.
Future Vision
Expand to multi-modal inputs (screenshots/videos), integrate with productivity tools, and add collaborative features.