Selected Projects
Pi0.5 Vision-Language-Action Integration
Full end-to-end implementation of Pi0.5 VLA policy in ark_ml framework
Complete integration featuring dataset loading, model definition, training pipelines, evaluation utilities, and benchmarking. Includes full Pi0.5 model architecture with vision encoder, Q-Former multimodal fusion, and FAST-token pipeline for continuous action prediction.
RubikNet
Deep learning solution for solving Rubik's Cube with neural networks
A computer vision and reinforcement learning project that combines image recognition with algorithmic problem solving to solve the Rubik's Cube autonomously.
Brax training viewer
Developed a reinforcement learning pipeline using Brax and MuJoCo to train and visualize PPO policies for robotic control.
multi-agent simulation with dynamic environment replication and real-time policy visualization. Leveraged JAX for accelerated training across 4096 parallel environments with efficient inference deployment.
Gym-master
Reinforcement Learning algorithms for solving OpenAI Gym environments
Implementation of various RL algorithms using NumPy to understand fundamentals without abstractions. Focus on policy gradient and Q-learning methods.
KDAIL-forklift
Advanced SLAM robot with UWB-based localization
ROS2-based autonomous forklift robot using RPLIDAR A2-M8 for mapping and UWB sensors for precise localization, replacing traditional encoders for improved accuracy.
Maze Solver Bot
ESP-32 powered autonomous maze-solving robot
Embedded system using PID controllers and left-first search algorithm for efficient maze navigation. Demonstrates real-time pathfinding and control systems.
Pull Requests & Contributions
sktime Time Series Library
Pull Request: Add LLM prompts section with automated template sync
Developed documentation system for sktime featuring LLM prompts and automated Sphinx extensions. Created template-include directive that dynamically pulls content from code, ensuring docs stay synchronized with estimator templates.
pgmpy Probabilistic Graphical Models
Pull Request: Improvements to CI workflow
Built PR-focused flake8 workflow that checks only changed .py files using git diff. Supports incremental code quality adoption without disrupting existing codebases, with plans for expanded linting features.