> I implement neural networks from the math up — backprop and optimizers in NumPy, inference in C++ compiled to WebAssembly.
B.Tech AI & ML · Amity University Bangalore · Class of 2028
I learn ML by building it without the safety net — and then checking whether it actually works. I wrote a neural network from scratch in C++ and compiled it to WebAssembly, so it runs in your browser, right below. Then I rebuilt the same architecture in NumPy, with no autograd, to compare SGD, Momentum and Adam by hand (98.07% on MNIST). When I audited my fake-news classifier, I found 18% of its test set had leaked in from training; the honest number is lower, and it's the one on this site.
Handwritten digit recognizer — backprop, matrix library and layers written from scratch in C++, compiled to WebAssembly. Draw below and watch the hidden layer respond.
Draw a digit (0–9)
Fully connected neural network built with NumPy only, comparing SGD, Momentum, and Adam optimizers.
TF-IDF + logistic regression on full articles (title + body). I removed 8,577 duplicate articles that were leaking into the test set, then tested what the model relies on: its top-weighted features were source and formatting cues, but stripping 38 of them costs only 1.7 points.
XGBoost rockfall-risk classifier trained on a 100,000-row synthetic slope-stability dataset, served through FastAPI to a React dashboard with runtime-adjustable risk thresholds. SMS alerts via Twilio fire on critical risk.
Open to applied ML internship conversations. Email is fastest: jadenisaac10@gmail.com