Intelli-Credit — AI-Powered Corporate Credit Appraisal Engine

Hackathon Solution for the "Intelli-Credit" Challenge
Theme: Next-Gen Corporate Credit Appraisal: Bridging the Intelligence Gap

---

🏗 Architecture Overview

intelli-credit/
├── backend/               # FastAPI + Python AI backend
│   ├── main.py            # App entry point
│   ├── config.py          # API keys & settings
│   ├── requirements.txt
│   ├── .env.example       # Copy → .env with your keys
│   ├── routers/
│   │   ├── health.py      # GET /api/health
│   │   ├── ingestor.py    # POST /api/ingestor/upload
│   │   ├── research.py    # POST /api/research/run
│   │   └── recommendation.py  # POST /api/recommendation/generate
│   └── services/
│       ├── pdf_parser.py       # pdfplumber + PyPDF2 cascade
│       ├── gst_analyser.py     # GST vs Bank cross-check
│       ├── gemini_service.py   # Google Gemini AI calls
│       ├── web_search.py       # Serper.dev web search
│       ├── scoring_engine.py   # Five-Cs scoring model
│       └── cam_generator.py    # ReportLab PDF generation
└── frontend/              # React + Vite frontend
    ├── src/
    │   ├── pages/
    │   │   ├── HomePage.jsx        # Landing / dashboard
    │   │   ├── IngestorPage.jsx    # Step 1: Document upload
    │   │   ├── ResearchPage.jsx    # Step 2: Research agent
    │   │   ├── RecommendationPage.jsx  # Step 3: Credit decision
    │   │   └── CAMPage.jsx         # Step 4: CAM download
    │   ├── components/
    │   │   └── Layout.jsx          # Sidebar + toast system
    │   ├── context/
    │   │   └── AppContext.jsx      # Global session state
    │   └── api.js                  # Axios API client
    └── vite.config.js              # Proxy → localhost:8000

---

🔑 API Keys Required

Service	Purpose	Get It
Google Gemini	LLM for extraction, research synthesis, credit decisions	aistudio.google.com — Free tier available
Serper.dev	Web search for news, litigation, regulatory research	serper.dev — 2500 free queries/month
OpenAI (optional)	Fallback LLM if Gemini unavailable	platform.openai.com

---

🚀 Quick Start

Step 1 — Clone & set up API keys

cd backend
copy .env.example .env

Edit backend/.env:

GEMINI_API_KEY=AIza...your_key_here
SERPER_API_KEY=your_serper_key_here

Step 2 — Backend setup

cd backend

# Create virtual environment
python -m venv venv

# Activate (Windows)
venv\Scripts\activate

# Activate (Mac/Linux)
source venv/bin/activate

# Install dependencies
pip install -r requirements.txt

# Start FastAPI server
python main.py

Backend runs at: http://localhost:8000
API docs (Swagger): http://localhost:8000/docs

Step 3 — Frontend setup

cd frontend

# Install dependencies (if not already done)
npm install

# Start dev server
npm run dev

Frontend runs at: http://localhost:5173

---

📋 Three-Step Workflow

🗂 Step 1: Data Ingestor

• Upload PDF documents (annual reports, financial statements, legal notices, MCA filings, rating agency reports)
• Optionally provide GST JSON (monthly_turnover, gstr_3b_tax_paid, gstr_2a_itc_claimed) and Bank Statement JSON (monthly_credits) for circular trading detection
• AI extracts: Revenue, EBITDA, PAT, Debt, Net Worth, DSCR, ICR, D/E ratio, and more
• Engine flags circular trading if GST vs bank discrepancy > 30%

GST JSON format:

{
  "monthly_turnover": { "2024-01": 5000000, "2024-02": 5200000 },
  "gstr_3b_tax_paid": { "2024-01": 900000, "2024-02": 936000 },
  "gstr_2a_itc_claimed": { "2024-01": 750000, "2024-02": 780000 }
}

Bank Statement JSON format:

{
  "monthly_credits": { "2024-01": 5100000, "2024-02": 5400000 }
}

🔍 Step 2: Research Agent

• Enter company name, sector, and key promoter names
• Engine searches news, litigation (NCLT/DRT/SEBI), and regulatory updates via Serper.dev
• Input qualitative due diligence notes — AI adjusts the credit score (−30 to +10 points) based on field observations
• Returns early warning signals and a synthesized research brief

📊 Step 3: Recommendation & CAM

• Enter loan details (amount, purpose, tenure, collateral)
• Engine runs Five-Cs scoring (transparent weights):
  • Character 25% — Promoter integrity, governance, litigation
  • Capacity 30% — DSCR, ICR, EBITDA margin, operating cash flows
  • Capital 20% — Net worth, D/E ratio, leverage
  • Collateral 15% — Security cover, collateral quality
  • Conditions 10% — Sector outlook, macro environment
• Gemini generates: Decision (APPROVE / CONDITIONAL_APPROVE / REJECT), recommended amount, interest rate, risk premium (bps)
• Download the Credit Appraisal Memo (CAM) PDF

---

🏦 Indian Context Features

Feature	Detail
GSTR-2A vs 3B	Flags ITC ≥95% of tax paid (circular trading signal)
CIBIL/Rating context	Prompts include India credit bureau references
NCLT/DRT/SEBI	Litigation search targets India-specific courts
eCourt portal	Included in search queries
RBI regulations	Sector regulatory searches include RBI circulars
INR Crores	All financials in Indian units
MCLR/Repo	Interest rate framing in India banking context

---

🔌 API Reference

Endpoint	Method	Description
/api/health	GET	Check API keys, backend status
/api/ingestor/upload	POST	Upload PDFs + optional GST/bank JSON
/api/ingestor/session/{id}	GET	Retrieve extracted session data
/api/research/run	POST	Run web research agent
/api/recommendation/generate	POST	Generate Five-Cs score + decision
/api/recommendation/cam/{id}	GET	Download CAM PDF
/api/recommendation/session/{id}	GET	Get saved recommendation

Full Swagger UI: http://localhost:8000/docs

---

🛠 Troubleshooting

Backend won't start

• Check Python ≥ 3.9: python --version
• Activate virtual environment before running
• Install deps: pip install -r requirements.txt

"AI Offline" in sidebar

• Add GEMINI_API_KEY to backend/.env
• Restart the backend after editing .env

No search results

• Add SERPER_API_KEY to backend/.env
• Demo mode: without the key, placeholder results are returned

PDF extraction returns empty financials

• Complex scanned PDFs may need OCR (not included in base version)
• Try uploading a text-native PDF first
• pdfplumber handles most Indian CA-certified financial statements

CAM download fails with 404

• Run /api/recommendation/generate first
• CAM is generated as part of the recommendation step

---

📐 Scoring Model Reference

Total Score = Σ (C_score × weight)

Decision Logic:
  ≥ 65 → APPROVE
  50–64 → CONDITIONAL_APPROVE
  < 50 → REJECT

Grade Table:
  85+ → AAA    75–84 → AA    65–74 → A
  55–64 → BBB  45–54 → BB    35–44 → B    <35 → C

---

🧑‍💻 Tech Stack

Layer	Technology
Backend	FastAPI, Python 3.11
LLM	Google Gemini 1.5 Flash
Web Search	Serper.dev (Google Search API)
PDF Parsing	pdfplumber + PyPDF2
PDF Generation	ReportLab
Frontend	React 18 + Vite 7
Charts	Recharts (Radar + Bar)
Icons	Lucide React
State	React Context API

---

Built for the Intelli-Credit Hackathon · India-aware AI Credit Engine