Urdu Voice-to-Voice RAG Agent

Problem

Most voice-AI tooling is built and tuned for English first, leaving Urdu speakers underserved by conversational assistants.

Solution

Built an end-to-end Urdu voice-to-voice conversational agent using open-source STT, LLM, and TTS models from Hugging Face, with a RAG pipeline for context-aware responses. Deployed inference on Runpod GPUs with the frontend hosted on Vercel.

Tech Stack

• Whisper (Hugging Face) — Open-source speech-to-text with usable Urdu transcription.
• RAG pipeline — Grounds responses in retrieved context instead of relying on the LLM alone.
• Runpod — Cost-efficient GPU inference for the STT/LLM/TTS pipeline.
• Vercel — Hosts the frontend interface for the assistant.

Challenges

Write-up coming soon.

Impact

Metrics coming soon.

Lessons Learned

Write-up coming soon.