PlacementIQ

AI-powered placement risk modelling for education loan portfolios.

PlacementIQ is a lender-facing prototype for NBFC education loan monitoring. It predicts whether a borrower is likely to be placed before or soon after moratorium exit, then gives risk teams a forward-looking Placement Risk Score, placement probabilities, expected salary band, explainability drivers, and recommended interventions.

Requirements

• Python: 3.11.0 or above
• Node.js: 24 or above
• NPM: Latest version

Project Structure

main/                      Next.js frontend dashboard
ml/                        Python backend and ML models
ml/api.py                  FastAPI scoring API
ml/scripts/                Data generation and training scripts
ml/models/                 Trained model artifacts
ml/data/                   Seed and synthetic data
requirements.txt           Python dependencies
package.json               Frontend dependencies (in main/)
Dockerfile                 API container setup

Setup & Running

1. Backend (FastAPI)

Create a virtual environment and install dependencies:

python -m venv .venv
.\.venv\Scripts\activate
pip install -r requirements.txt

Run the scoring API:

python -m uvicorn ml.api:app --reload

The API will be available at http://127.0.0.1:8000/docs.

2. Frontend (Next.js)

Navigate to the main folder and install dependencies:

cd main
npm install

Run the development server:

npm run dev

The dashboard will be available at http://localhost:3000.

---

Train The Models

Generate the synthetic training dataset:

python ml/scripts/generate_synthetic_data.py --rows 25000

Train the models:

python ml/scripts/train_boosted_models.py

Current Metrics

Trained on 25,000 synthetic rows with a 20,000 / 5,000 train-test split:

• 3-month placement, Stage 2 LightGBM: AUC 0.8122, accuracy 0.7462
• 6-month placement, Stage 2 LightGBM: AUC 0.8676, accuracy 0.8812
• 12-month placement, Stage 2 LightGBM: AUC 0.8967, accuracy 0.9542
• Salary LightGBM: MAE 0.9625 LPA, RMSE 1.2041 LPA

Data Strategy

PlacementIQ uses a hybrid data approach to ensure both realism and privacy:

• Institutional Data (Actual): The model is seeded with actual NIRF Engineering 2024 rankings. This provides a grounded baseline for institutional quality, placement history, and regional demand signals.
• Student-Level Factors (Synthetic): Individual student attributes—such as academic performance (CGPA), internship history, certifications, and interview activity—are synthetically generated. This approach allows for a robust, diverse dataset without compromising sensitive private borrower information.

Normalization & Bias Mitigation

• GPA Scaling: Academic results are normalized to a standard 10-point scale, accommodating both Indian and international grading systems.
• Fairness by Design: Sensitive demographic features (Gender, Caste, Religion, Origin) are excluded from the model features to prevent automated bias, used only for auditability and fairness monitoring.

Responsible AI

PlacementIQ is decision support, not an automated loan rejection tool. Every model score is backed by SHAP explainability artifacts so a lender can inspect the drivers behind the score.

Docker

docker build -t placementiq-api .
docker run -p 8000:8000 placementiq-api