Intelligence built
for a fashion boutique

Tracks posts and reels across two boutique Instagram accounts — likes, views, comments, and engagement trends. A custom Node.js importer feeds a Supabase database; a stored procedure aggregates stats per boutique. The dashboard surfaces top-performing content and weekly patterns in a clean, filterable view.

A Kotlin foreground service that watches the device's call recordings directory and silently uploads every new .m4a file to Supabase Storage. Handles Android 11–15 storage permission models, survives reboots via a BroadcastReceiver, and self-restarts every 5 hours to work around the Android 15 FGS time limit. Upload status is persisted in SharedPreferences and shown live in the app UI.

A 3-node LangGraph pipeline that runs automatically on every new recording.

Whisper transcribes and lingua-py tags each segment as Malayalam, English, or Manglish. Haiku normalises the code-switched text and infers call direction (incoming / outgoing) from the opening lines. Sonnet produces a plain-English summary, structured action items, sentiment score, and language classification — stored as JSON in Supabase.

A web interface listing every processed call — caller name, direction badge, sentiment, and duration. Each row links to a detail view with the full diarised transcript (speaker labels + timestamps), audio playback streamed from Supabase Storage, the AI-generated plain-English summary, and extracted action items. Processing status (pending → processing → done / failed) is tracked per file with automatic retry for stale jobs.