Application

Van Saathi

Indic-language voice agent — multi-speaker transcription and diarisation fused with geospatial field data, answered back as speech.

Launch

Speakers diarised

Speech

Output

WhisperSarvam STT / TTSpyannoteClaudeECS Fargate

DomainsForest & Governance

Voice agent embedded in a field data platform. Long meeting recordings are transcribed in local languages (Whisper + Sarvam STT) and split by speaker with pyannote diarisation (up to 20 speakers, with speaker enrolment). Transcripts are fused with boundary-mapping and inventory data so Claude can draft plans grounded in both, and results are returned through Sarvam TTS — no literacy required. The API runs as a containerised service on AWS ECS Fargate.

full case study in progress — architecture, dataset notes, and evaluation details to follow