🚦 Scones Unlimited: Image Classification Workflow on AWS

📌 Project Overview

Scones Unlimited is a serverless, event-driven workflow for image classification of delivery vehicles. It uses AWS Step Functions, Lambda, and a deployed machine learning model to classify incoming images (e.g., bicycles, motorcycles).
This project demonstrates how to build a scalable ML pipeline that ingests input images, classifies them using an inference endpoint, and routes results accordingly.

---

🛠️ Tech Stack

• AWS Lambda – serverless compute functions
• AWS Step Functions – workflow orchestration
• Amazon SageMaker – ML model endpoint for inference
• Amazon S3 – input/output data storage
• Amazon SNS (optional) – notifications on errors or events
• Python – Lambda function code

---

⚙️ Workflow Architecture

1. Image Ingestion – Input images are uploaded as Base64 strings.
2. Preprocessing Lambda – Decodes and prepares the image for inference.
3. Inference Lambda – Sends image to SageMaker endpoint for prediction.
4. Postprocessing Lambda – Interprets predictions and decides next steps.
5. Fan-out Support (Optional) – Workflow can process multiple images in parallel.
6. Visualization – Model predictions are visualized with test data.

---

✨ Features

• ✅ Event-driven architecture with AWS Step Functions
• ✅ ML inference using SageMaker deployed model
• ✅ Parallel fan-out workflow (optional challenge)
• ✅ Data visualization for predictions (extra)
• ✅ Error handling with Step Function + SNS notifications (optional)
• ✅ Scalable, serverless deployment

---

📊 Visualization Example

Here’s an example of the model output visualization with CIFAR-10 vehicle classes:

---

🚀 Getting Started

Prerequisites

• AWS account with access to Lambda, Step Functions, S3, SageMaker
• Python 3.8+
• IAM roles configured for Lambda & Step Functions

Setup

1. Deploy Lambda functions (lambda_function.py)
2. Create SageMaker endpoint with trained CIFAR model
3. Define Step Function using state_machine.json
4. Test with sample images

---

📂 Project Structure

├── lambda.py
├── CIFARStateMachine.asl.json
├── visualization_screenshots/
│   └── [visualization images]
├── starter.ipynb
├── test.lst
├── train.lst
└── README.md

---

📝 Results & Learnings

Through building this project, I was able to:

• Learn how to design event-driven architectures with AWS Step Functions.
• Understand how to chain multiple Lambda functions into a production-style workflow.
• Deploy and query a SageMaker inference endpoint for real-time ML predictions.
• Visualize model outputs to monitor confidence and ensure threshold compliance.
• Explore optional challenges like parallel fan-out workflows, error notifications with SNS, and scaling the workflow with more classes.

Key Takeaway:
This project reinforced the importance of combining machine learning with cloud-native serverless design to create systems that are not only accurate, but also scalable, maintainable, and production-ready.

---

🔮 Optional Enhancements

• Add more classes from CIFAR dataset (cars, trucks, airplanes, etc.)
• Build a dummy/test data generator for continuous image input
• Use Step Functions fan-out for parallel image processing
• Add SNS topic for error notifications

---

📸 Demo Screenshots

(Insert screenshots of your Step Function running successfully, Lambda logs, and predictions here)

---

📖 Acknowledgements

This project is part of the Udacity Machine Learning Engineer Nanodegree.
Built with ❤️ using AWS services.