---
title: SME Credit Risk AI
emoji: 🏦
colorFrom: indigo
colorTo: blue
sdk: streamlit
sdk_version: 1.55.0
app_file: app.py
pinned: true
license: mit
---

# 🏦 AI-Powered SME Credit Risk Assessment

**Final Project — AI Engineering Bootcamp Batch 10 | Author: 1na37**

> Explainable · Prescriptive · Multilingual Credit Intelligence

---

## ✨ Features

| Feature | Description |
|---------|-------------|
| 🤖 **Ensemble ML** | XGBoost + LightGBM + Random Forest (soft voting, 2:2:1 weights) |
| 🔍 **XAI Engine** | SHAP Waterfall Plot per applicant — *why* each decision |
| 💰 **Risk Metrics** | PD Score + Expected Loss (PD × LGD × EAD, Basel II) |
| 🎮 **What-If Sim** | Interactive slider simulation, real-time score update |
| 💬 **LLM Narrative** | Gemini → Groq → Static fallback (auto chain) |
| 🌍 **Multilingual** | Bahasa Indonesia / English / Hindi |

---

## 📊 Dataset

- **Source**: German Credit Dataset (UCI ML Repository, id=144)
- **Rows**: 1,000 | **Original features**: 20 | **After augmentation**: 28
- **Augmentation**: Synthetic SME-relevant features added via Python:
  - `digital_presence_score`, `has_social_media`, `ecommerce_monthly_volume`
  - `has_npwp`, `has_siup`, `business_age_years`, `monthly_cash_flow`, `num_employees`
- **Class balance**: SMOTE applied on training set only (no data leakage)

---

## 🏗️ Architecture

```
Input (28 features)
    ↓
One-Hot Encoding + StandardScaler + SMOTE (train only)
    ↓
┌─────────────────────────────────────┐
│  Soft Voting Ensemble (2:2:1)       │
│  ├── XGBoost  (n=300, lr=0.05)     │
│  ├── LightGBM (n=300, lr=0.05)     │
│  └── RandomForest (n=200)           │
└─────────────────────────────────────┘
    ↓
PD Score → EL = PD × LGD(40%) × EAD
    ↓
SHAP TreeExplainer → Waterfall Plot
    ↓
LLM Narrative (Gemini → Groq → Fallback)
```

---

## 📈 Evaluation

- **ROC-AUC** — primary metric (industry target: >0.75)
- **KS Statistic** — separation between good/bad credit distributions
- **F1-Score** — balanced precision/recall for imbalanced data
- **5-Fold Cross-Validation** — SMOTE inside ImbPipeline (no leakage)
"Model menunjukkan performa sempurna karena synthetic features dirancang berkorelasi kuat dengan target,
sebagai simulasi kondisi ideal. Dalam deployment nyata,
fitur-fitur ini akan diganti dengan data aktual dari sumber eksternal
seperti API e-commerce, data perpajakan, dan rekening koran."
![Evaluation Report](evalsme.png)
---

## 🚀 Run Locally

```bash
git clone https://huggingface.co/spaces/1na37/sme-credit-risk
cd sme-credit-risk
pip install -r requirements.txt

# Train model first (run train_colab.py in Google Colab)
# Download model_export.zip, extract to model/ folder

streamlit run app.py
```

---

## 🔑 API Keys (Optional)

In HF Spaces → **Settings → Repository Secrets**:
- `GEMINI_API_KEY` — [Google AI Studio](https://aistudio.google.com) (free tier)
- `GROQ_API_KEY` — [console.groq.com](https://console.groq.com) (free tier)

App works without API keys using static fallback template.

---

## 📚 References

- [German Credit Dataset — UCI ML Repository](https://archive.ics.uci.edu/dataset/144/statlog+german+credit+data)
- [SHAP Library](https://shap.readthedocs.io/)
- Basel II Credit Risk Framework (LGD = 40% standard)

---

*Built with ❤️ for  AI Engineering Bootcamp Batch 10*