Handwritten Nepali Character Recognition

Handwritten character recognition is a solved demo for Latin script — and a genuinely hard problem for Devanagari, where characters share a connecting headline (dika), differ by subtle strokes, and handwriting styles vary enormously. For our final-year Computer Engineering project at IOE, my team of four took on recognizing handwritten Nepali characters.

We built a deep convolutional neural network in Keras, trained on a labeled dataset of handwritten Devanagari characters, and wrapped it in a Flask interface so anyone could try the model against their own handwriting rather than taking our accuracy numbers on faith.

We trained the same architecture under two optimizers — Adam and Nadam — and compared them properly instead of trusting defaults. Nadam, which adds Nesterov momentum to Adam's adaptive learning rates, consistently outperformed and carried the model to 99% validation accuracy.

As an undergraduate lesson, that comparison outlasted the accuracy number: the optimizer is a hyperparameter, and measuring beats assuming.

This project — built with Anish Baral, Diwash Khanal, and Prabin Baskota — is where my interest in machine learning started. Years later it led me back: my Master's work on neural network training failures is, in a real sense, the theory behind everything we debugged empirically here.