Controlled experiment comparing LoRA adapters trained on Zero123++-generated 3D-consistent synthetic views vs. 2D-augmented data for single-image object detection on Google Scanned Objects.

Sole-authored deep learning framework in Python/NumPy. Dynamic computation graph, reverse-mode autograd, Conv2D via im2col, multi-head attention, transformers, Adam/AdamW. Published on PyPI.

Hierarchical CNN framework identifying second-order dynamical system parameters from step response images. Three-stage pipeline: regime classification → geometric observable estimation → physics-based parameter recovery.

Complete redesign and development of otokon.org and ituro.otokon.org (now ituro.org). HTML/CSS/JS.

Custom 2D rigid body engine in C++ with Python bindings. Articulated agents via joint constraints. Designed for RL training integration.

Classical CV system in Python/OpenCV. ORB features, essential matrix decomposition, trajectory reconstruction. Foundation for SLAM in GPS-denied environments.

High-performance C++ backend for Python-trained neural networks. Eigen + PyBind11 zero-copy bridge.

Unsupervised LSTM autoencoder monitoring multi-variate sensor data for flight anomaly detection.