End-to-End Machine Learning Deployment — Lightweight, Containerized, Production-Ready ML Service

This project demonstrates how to convert a trained machine learning model into a fully deployable inference service using Docker and AWS EC2, optimized for low-resource cloud environments.

Unlike typical ML repositories that mix training and inference together, MDMLP (Minimum Deployable Machine Learning Project) intentionally separates them:

• Training happens offline (local / Colab / GPU)
• Only artifacts are deployed (preprocessor + model pickle)
• The Docker image contains inference-only dependencies
• This keeps the container small, fast, and cloud-efficient

This is the same deployment philosophy used in real-world ML production systems.

Problem Statement

The model predicts student performance (final math score) from demographic and academic features such as:

gender
race_ethnicity
parental_level_of_education
lunch
test_preparation_course
writing_score
reading_score

Dataset used:
https://www.kaggle.com/datasets/spscientist/students-performance-in-exams

Model Selection Summary

Multiple regression models were trained and evaluated. Ridge Regression and Linear Regression scored almost identically, but Linear Regression had a smaller model size and faster inference time.

Final deployed model: Linear Regression
Reason: Similar accuracy to Ridge Regression, but smaller storage size and more efficient inference.

Top test-set performance scores:

Linear Regression — R2: 0.8803
Ridge Regression — R2: 0.8806
CatBoost — R2: 0.8516
AdaBoost — R2: 0.8488
Random Forest — R2: 0.8473
XGBoost — R2: 0.8278
Lasso — R2: 0.8253
KNN — R2: 0.7838
Decision Tree — R2: 0.7206

API Inputs

The app accepts 7 user inputs (numeric or dropdown fields):

gender
race_ethnicity
parental_level_of_education
lunch
test_preparation_course
writing_score
reading_score

The output is a predicted math score.

Architecture (Training → Artifacts → Deployment)

Training happens offline
↓
Export serialized artifacts (model.pkl + preprocessor.pkl)
↓
Build Docker image with inference-only environment
↓
Run container on AWS EC2
↓
User inputs features via web form
↓
Model returns score prediction

Key design decisions:

• Training code is not packaged inside deployment container
• Only serialized artifacts are included
• Container stays lightweight and boots quickly even on free-tier EC2

Cost Optimization (AWS Free Tier)

AWS Free Tier has a 720-hour/month limit.
To avoid unnecessary billing, the EC2 instance is intentionally shut down when not in use.
The container remains intact and can be restarted any time.

This reflects real-world cost-aware ML deployment practices used in startups.

Docker Usage

Build image (optional)

docker build -t student-performance .

Run container locally

docker run -d -p 8080:8080 student-performance

Open in browser:

http://localhost:8080

Deploy on AWS EC2 (Full Workflow)

1. SSH into EC2

ssh -i <your-key.pem> ubuntu@<your-ec2-public-ip>

2. List Docker images

docker images

3. Run container

docker run -d -p 8080:8080 <docker-image-name>

4. Confirm the container is running

docker ps

5. Access application from browser

http://<your-ec2-public-ip>:8080

6. Stop / restart container

docker stop <container-name>  
docker start <container-name>

IMPORTANT
Ensure port 8080 is open in EC2 security group.

Training Repository Link

This deployment repository intentionally contains inference-only code.

The full training pipeline including:

• model comparisons
• evaluation metrics
• hyperparameter tuning
• artifact generation (model.pkl and preprocessor.pkl)

…is available in the companion repository:

https://github.com/fishyyuser/ML-Project

MDMLP uses the exported artifacts generated from that repository.

Tech Stack

Python
Flask
Scikit-Learn
Docker
AWS EC2
HTML / CSS for simple UI

License

MIT