DAVID
GONZALEZ

Project Overview

Production-grade machine learning system that can accurately recognize handwritten digits in real-world conditions with 98.63% accuracy. Built from scratch using 21,000+ European handwritten digit images, this end-to-end project demonstrates professional ML engineering, from data preprocessing and CNN training to REST API deployment and interactive web interface. Unlike typical academic exercises, this system is containerized, production-ready, and serves real-time predictions in under 100ms.

Challenge

The challenge wasn't just achieving high accuracy, it was engineering a complete production system. Specific obstacles included: (1) transforming 21,000+ heterogeneous Swiss handwritten digit images into a clean, MNIST-compatible dataset through robust preprocessing, (2) eliminating train-test mismatch by ensuring inference preprocessing perfectly replicates training preprocessing, (3) building a production-ready API, and (4) containerizing the entire stack for reproducible deployment.

Approach

Data Engineering: Designed a 5-step preprocessing pipeline (grayscale → LANCZOS resize to 28x28 → color inversion → normalization) that transforms raw images into MNIST format.
Model Architecture: Built a 3-layer CNN with batch normalization, dropout regularization, and strategic callbacks (EarlyStopping, ReduceLROnPlateau, ModelCheckpoint)
Backend Development: Implemented FastAPI with dependency injection for model loading, Pydantic for request/response validation, environment-based configuration (dev/prod/test modes), and structured error handling.

Results & Impact

• 98.63% test accuracy on 21,000+ image dataset
• >100ms inference time per prediction
• lightweight and efficient
• Zero train-test mismatch through identical preprocessing pipelines