Welcome to our AI Model Pipeline. This project showcases the complete lifecycle of building, training, evaluating, and deploying a machine learning model using a modern MLOps workflow.
This project addresses the real-world challenge of manually verifying and assigning applicants to the correct sub-schemes under the National Social Assistance Program (NSAP)—a critical welfare initiative by the Government of India.
The system uses machine learning to predict the correct NSAP scheme for each applicant, reducing human error and accelerating aid distribution.
Key highlights:
- End-to-end machine learning pipeline design
- Automated data processing, model training, and evaluation
- Leaderboard-based model selection
- Seamless deployment for real-time or batch inference
The pipeline is built to be scalable, modular, and easily adaptable to other government schemes or datasets.
- Source: AI Kosh NSAP dataset
- Features: Age, gender, income level, disability status, marital status
- Labels: Scheme codes
- IGNOAPS (Old Age)
- IGNWPS (Widow)
- IGNDPS (Disability)
- Removed or handled missing/inconsistent values
- Categorical fields encoded into numerical format
- Data normalized where required
- Balanced dataset using under/over sampling techniques
- Model Used: Random Forest Classifier
- Platform: IBM Watsonx.ai
- Task: Multi-class classification
- Goal: Predict the most suitable NSAP sub-scheme based on applicant details
The training pipeline is fully automated and includes the following stages:
- Data preprocessing
- Model training
- Model evaluation and validation
- Leaderboard generation for ranking models
| Step | Screenshot |
|---|---|
| Training Initialization |  |
| Evaluation Started |  |
After training, all models are evaluated and ranked based on defined metrics (accuracy, F1-score, etc.). The leaderboard helps in selecting the top-performing model for deployment.
The top-performing model is deployed online via Watsonx.ai for real-time predictions on new applicant data.
The deployed model was tested to ensure real-world readiness. The system outputs confidence scores with each prediction, aiding human decision-makers.
Example results:
- IGNDPS – 70%
- IGNOAPS – 60%
- IGNWPS – 70%
- Accuracy: Verified through test set performance
- Confidence scores: Range from 50% to 70%
- Faster processing: Significantly reduces manual verification time
- Transparent results: Confidence scores enable better decision-making
- Python
- IBM Watsonx.ai
- GitHub Actions
- MLOps pipeline orchestration tools
- Fully automated and reproducible ML pipeline
- Model evaluation and ranking with leaderboard
- Fast, real-time deployment on IBM Watsonx.ai
- Accurate predictions with confidence scores
- Designed for scalability and reusability in other government schemes
Built with precision. Deployed with purpose.