How arijit-chandra/mlops-machine-learning-based-visa-certification-prediction works
Main Components
a. Frontend / API Layer
• app.py (Flask application entrypoint)
• demo.py (CLI or quick-demo script)
• templates/*.html & static/css (UI assets)
b. ML Backend / Business Logic
• us_visa_prediction package
• components/ (data_ingestion, data_transformation, validation, model_trainer, model_evaluation, model_pusher, model_factory)
• pipline/ (training_pipeline.py, prediction_pipeline.py)
• configuration/ (mongo_db_connection.py, azure_connection.py)
• data_access/ (easyvisa_data.py)
• entity/ (config_entity, artifact_entity, estimator entities)
• utils/main_utils.py, logger, exception
c. Data Stores & External Services
• MongoDB Atlas (metadata, schema registry)
• Azure Blob Storage (model artifacts)
• Kaggle-sourced dataset (CSV)
• Evidently AI (monitoring dashboards)
d. CI/CD & Deployment
• Dockerfile & .dockerignore
• GitHub Actions workflow (build → test → deploy)
• Azure (Container Registry or Web App, Azure Monitor)
Relationships & Data Flow
• User → Flask App: submits input or triggers training/prediction
• Flask App → Prediction Pipeline: loads latest model from Azure Blob via azure_storage client
• Prediction Pipeline → MongoDB: reads schema/config or logs prediction events
• Training Pipeline → Data Access: ingests EasyVisa.csv
• Training Pipeline → Components: validation → transformation → model_trainer → model_evaluation → model_pusher
• model_pusher → Azure Blob Storage: uploads new model
• CI/CD → Docker Build → Azure Deployment
• Evidently → Monitors model performance in production
Architectural Patterns & Design Principles
• Layered modular design (components, pipeline, data_access, configuration, entity)
• Separation of concerns (training vs. prediction pipelines)
• Externalized configuration (model.yaml, schema.yaml, environment variables for DB & storage)
• Infrastructure as code / containerization
• CI/CD automation with GitHub Actions
• MLOps best practices (model registry, monitoring, reproducibility)
Technologies & Frameworks
• Python, Flask
• CatBoost (used in notebooks, via model_factory)
• MongoDB Atlas (NoSQL metadata store)
• Azure Blob Storage (model artifact store)
• Docker, GitHub Actions, Azure CI/CD
• Evidently AI for drift monitoring
• Jupyter Notebooks for EDA & prototype training
Elements
• Explicit representation of config files (model.yaml, schema.yaml) feeding into configuration components
• A “Monitoring” block (Evidently AI, Azure Monitor) hooked to production logs