Mathematics & Computer Science · Building from first principles
I'm a maths and CS student who builds things from the ground up to understand how they really work — no PyTorch, no Unity, no shortcuts. Currently exploring relativistic physics and deep learning theory.
End-to-end pipeline from Monte Carlo physics simulation to live neural network inference. Classifies electrons, pions, muons, and gamma rays using energy deposition patterns across a sampling calorimeter.
- Stack: C++20 (Geant4 11) · Java 21 (Spring Boot) · Apache Kafka · PostgreSQL · Docker
- Model: 82% accuracy · 2,564 parameters · 3.2 KB · no ML libraries — built from first principles
- Physics-informed feature engineering (layer ratios, shower depth, hadronic punch-through analysis)
A fully connected MLP built in plain Java — SGD, backpropagation, binary model persistence — using nothing but standard Java collections. Can be tested live on - https://tensorless.onrender.com
- Architecture:
784 → 128 → 64 → 10· 85–95% accuracy on MNIST - A fully functional website which lets users test the model and build a community that provides additional training data to improve the model.
- No external libraries of any kind
An interactive OpenGL visualiser exploring the action of Möbius transformations on the Riemann Sphere, built as a companion to the research paper "Analytical Correspondence between Möbius Transformation and Relativistic Physics".
- Special & general relativity
- Deep learning theory
- Game engine architecture
"Building a neural network from scratch teaches you more than using PyTorch ever could."
