Invisible Banker — HACK MUJ 4.0

The Invisible Banker: Turning Financial Intent into Action

An AI financial assistant that understands Indian language voice/text, interprets financial documents, and guides users through verified financial product workflows with complete source traceability.

Architecture

USER VOICE/TEXT → SARVAM STT → GROQ LLM → SUPABASE → RAG KNOWLEDGE → ELIGIBILITY ENGINE → DASHBOARD

Stack

Layer	Technology
Frontend	React 18 + Vite + Tailwind CSS
Backend	Python 3.11 + FastAPI + Pydantic
Database	Supabase (PostgreSQL + pgvector)
LLM (Primary)	Groq (llama-3.3-70b-versatile)
LLM (Fallback)	Gemini 1.5 Flash
Voice STT	Sarvam AI (saaras:v4)
PDF Extraction	PyMuPDF + pdfplumber
Knowledge Ingestion	Firecrawl
Embeddings	sentence-transformers (multilingual)

Quick Start

Prerequisites

• Python 3.11+
• Node.js 18+
• Supabase account (project configured)

Backend Setup

cd backend
python -m venv venv
# Windows:
venv\Scripts\activate
# Mac/Linux:
source venv/bin/activate

pip install -r requirements.txt
cp ../.env.example .env
# Edit .env with your credentials

uvicorn app.main:app --reload --port 8000

Frontend Setup

cd frontend
npm install
cp .env.example .env.local
# Edit .env.local with Supabase public keys

npm run dev

Database Setup

cd backend
python scripts/init_db.py     # Create tables
python scripts/seed_products.py  # Seed financial products
python scripts/ingest_knowledge.py  # Ingest source documents

Project Structure

InvisibleBanker/
├── backend/
│   ├── app/
│   │   ├── main.py              # FastAPI app entry point
│   │   ├── core/                # Config, DB, security
│   │   ├── api/routes/          # API route handlers
│   │   ├── services/            # Business logic
│   │   │   ├── llm/             # LLM provider abstraction
│   │   │   ├── document/        # Document extraction
│   │   │   ├── knowledge/       # RAG pipeline
│   │   │   └── voice/           # Sarvam STT
│   │   ├── engine/              # Deterministic eligibility engine
│   │   └── models/              # Pydantic schemas
│   ├── scripts/                 # Setup, seeding, ingestion
│   ├── tests/                   # Unit tests
│   └── requirements.txt
├── frontend/
│   ├── src/
│   │   ├── pages/               # Route pages
│   │   ├── components/          # UI components
│   │   ├── services/            # API calls
│   │   ├── hooks/               # React hooks
│   │   └── lib/                 # Utilities
│   ├── package.json
│   └── vite.config.js
├── obsidian-vault/              # Project memory (not source code)
├── .env.example                 # Credential template (safe to commit)
├── .gitignore
└── README.md

Core Principles

• LLM understands. Sources provide evidence. Code decides. Database remembers. Human handles uncertainty. LLM explains.
• LLM never determines eligibility — deterministic Python engine does
• Every consequential claim traces to an official source
• Voice failure → text fallback (always works)
• Low confidence → NEEDS_REVIEW (never silent assumption)

Demo Scenarios

Case	Profile	Expected Result
Ravi — Shopkeeper	Age 32, Rajasthan, Business 3yr, Income ₹40k/mo, Loan ₹5L	MUDRA Tarun: ELIGIBLE
Priya — New Vendor	Age 19, Street vendor, No bank account	PM SVANidhi: NEEDS_REVIEW
Kumar — Defaulted	Age 45, Existing NPA	All products: INELIGIBLE

Environment Variables

See .env.example for all required variables.

CRITICAL: SUPABASE_SERVICE_ROLE_KEY is backend-only. Never expose to frontend.