Featured Projects
Hierarchical Vision-Based Identification of Second-Order Dynamical Systems from Step Response ImagesFeatured
This work presents a hierarchical deep learning framework for identifying the parameters of second-order dy namical systems directly from rendered step response images. Instead of performing direct end-to-end regression of the damp ing ratio and natural frequency—which often suffers from high variance and poor physical consistency—the proposed method decomposes the inverse problem into three structured stages. These stages include regime classification, geometric observable estimation, and physics-based parameter recovery through closed-form relations. By embedding classical system identification theory into the inference pipeline, this decomposition tightly constrains the hypothesis space and significantly improves generalization across heterogeneous response geometries
Physics-Based Reinforcement Learning Environment: rigidRLFeatured
A custom 2D rigid body mechanics engine built in C++ with direct Python bindings. Designed to simplify the integration of physics simulations into reinforcement learning workflows. It supports articulated agents via joint constraints and provides a transparent API for creating environments, managing state observations, and bridging the gap between low level physics implementation and high level training logic.
A Deep Learning Framework from Scratch: forgeNN Featured
Sole author of a deep learning framework developed in Python/NumPy, published on PyPI (pip install forgeNN). The framework features a dynamic computation graph with reverse-mode automatic differentiation, highly vectorized tensor operations, and a full suite of modules for building modern neural networks including transformers.
FPGA Accelerated Neural Predictive Control on Zynq SoC
Developing a high-performance control system implementing Neural Predictive Control (NPC) on Xilinx Zynq-7000 SoC. Offloading neural network inference to FPGA logic to achieve microsecond-level latency in real-time control loops. Integrating C++ control logic with hardware-accelerated components to optimize power and performance.
OTOKON and ITURO Website Remake
Redesigned and developed the official websites for OTOKON (otokon.org) and ITURO (ituro.org) using HTML, CSS, and JavaScript. The new designs focus on improved user experience, mobile responsiveness, and modern aesthetics to better represent the organizations' missions and activities.
Monocular Visual Odometry Pipeline
A classical computer vision system built from scratch in Python and OpenCV to estimate a camera's motion and reconstruct its trajectory from a sequence of images, a foundational component for SLAM in GPS-denied environments.
C++ Accelerated Inference Engine
A high-performance C++ backend for a Python-trained neural network. This project bridges the gap between Python prototyping and the speed required for real-time applications by moving core matrix operations to a release-optimized C++ build.
UAV Telemetry Anomaly Detection
An unsupervised deep learning system to monitor multi-variate sensor data in real-time and flag critical flight anomalies. The system uses a specialist LSTM Autoencoder trained on normal flight data to learn a signature of healthy system behavior.