Mobile app + AI

Hikam Shughro

A Flutter app with its own Node.js backend for a community organisation: finance, dues, attendance, meetings, minutes, procurement, live sessions and a book library. Maktabah AI answers from reference books with citations (RAG on a Qdrant vector database), and meeting recordings are transcribed with Whisper large-v3 into minutes.

Hikam Shughro landing page with the organisation management mobile app

The problem

Coordination ran on handwritten notes and WhatsApp groups where things got lost: cash, dues, attendance, minutes and procurement.

What it does

  • One mobile app with 22 feature areas, including finance, dues, attendance, meetings, minutes, procurement, policies, live sessions, prayer times and a book library.
  • Passwordless login with a one-time code sent over WhatsApp, and permissions per position.
  • Maktabah AI answers questions from the reference books and cites the passages it used.
  • Meeting recordings are transcribed with Whisper large-v3 and turned into minutes.
  • Live sessions through LiveKit, dues and donations through Midtrans, push reminders through Firebase.

How it is built

  • Flutter app (Riverpod) with its own self-hosted backend: Fastify, MySQL through Prisma, Redis with BullMQ and Socket.IO.
  • A Qdrant vector database for the book assistant.
  • PM2 process configuration; released on Google Play.

How AI is used

  • RAG: documents are split into chunks, embedded, stored in Qdrant, retrieved and passed to an OpenAI-compatible chat model, which answers with citations.
  • Speech-to-text with whisper-large-v3 through a self-hosted proxy, then an AI step that writes the minutes.
  • A vision model can read scanned book pages. Model names are set by environment.

Honest notes

  • Automated test coverage is minimal.
  • The same codebase also served as the source for a second organisation’s app that has not been released.

Engineering depth

Counts taken from the repository itself, not performance claims.

22
feature areas in the app
50
database models
~180
route handlers in 15 groups
107
commits