PORTFOLIO
Game Design Projects
Machine Learning Projects
Miscellaneous
Stellarium: A Space Odyssey | CAVE2 Application
Stellarium is a virtual reality project developed in Unity for the CAVE 2 system, It features over 107,000 stars and various constellations accurately placed as seen from Earth. Users can navigate space, explore constellations, and observe stellar movements over time, adjusted by user input.
Neon-Bites | PC Game
.png)
A thrilling cyberpunk food delivery game where players navigate a neon-lit city, avoiding obstacles and enemies to deliver orders on time while managing resources and upgrading their character. Features include dynamic driving mechanics, minimaps, and interactive NPCs.
SnAIder-Cut: XR VR Preproduction tool
Won Best Location AR at MIT XR Reality Hackathon 2024 by using Mixed Reality and Generative AI to visually generate and modify movie scenes in real-time, ensuring seamless collaboration and reducing filming delays. A pre-production tool for scene blocking and planning.
Sign Smash | FPS Mobile Game
.png)
Developed in Unity 3D, Sign Smash is an action-packed FPS shooter game featuring basic AI enemies, traps, tricks, and a challenging final boss with multiple paths for attack and defense. The game includes a revival system supported by ad rewards, enhancing player engagement and providing a dynamic and immersive mobile gaming experience.
Equity Research Project | Unreal Engine Application
Developed an Unreal Engine 5 application for UIC AHS to support equity research in the medical field, featuring a dialogue tree and utilizing MetaHuman and Nvidia Omniverse. Successfully distributed to research participants and medical professionals.
.png)
.png)
KILL THE MOTHERBOARD | Unity Multiplayer Game
Developed a 3-player Unity game for "cats & computer architecture" theme. Players cooperatively overheat the CPU by delivering a power surge or stopping the fan. This educational game teaches players about motherboard function through an engaging challenge.
Pixel Punks | Collaborative Pixel Art Website
At Solana Hacker House Bengaluru 2022, I contributed to a collaborative pixel art project on the Solana blockchain. Users could collectively create a piece that would be minted as an NFT. Each pixel change involved a small Solana transaction, seamlessly integrated within the website. This concept project explored user-driven art creation and ownership on the blockchain, though functionality was limited.
Auto-prompting for PaintSeg | Research project
Developed an innovative autoprompting system for PaintSeg, automating input generation for training-free object segmentation. Leveraged k-means clustering for color-based segmentation and the Dense Prediction Transformer (DPT) model to extract depth maps, creating precise binary and bounding box masks without manual input. Experiments on the DUTS dataset showed IOU scores between 45% and 55%, with enhancements up to 60% using a hybrid prompting strategy. This approach significantly streamlines the segmentation process and paves the way for further automation in image processing tasks.
Back
to top
Microscopy Segmentation and Object Detection
Worked on segmenting a 5x5x5 um section of the CA1 hippocampus using an Electron Microscopy Dataset, aiming for optimal Intersection over Union (IoU) scores. Implemented various techniques, starting with basic histogram segmentation using Otsu’s threshold, and progressing to deep learning with UNet, both from scratch and with a pre-trained VGG19 backbone. Developed object detection models, including Single Shot Detection (SSD) and Fast R-CNN with multiple backbones such as MobileNetV2 and DenseNet121. Achieved significant improvements in precision, recall, and IoU scores, demonstrating robust detection and classification capabilities for complex visual data.
Back
to top
Azure Virtual Avatar Project (In Development)
Implemented Azure's Text-to-Speech model on our lab website to create a real-time talking avatar, leveraging Azure and OpenAI models for enhanced interactivity. Modified JavaScript code to integrate OpenAI for conversational capabilities, beyond Azure's basic text-to-speech. The integration involved setting up a server with Python scripts to connect the APIs, enabling dynamic conversations with the avatar. Additionally, we recorded and utilized a custom avatar with support from Microsoft team, enhancing the website's research functionalities. The project is ongoing, with further improvements underway.
Optiver Trading Challenge: Market Volatility Prediction
Utilized various regression models, including Linear Regression, SGD Regressor, Random Forest Regressor, and LGBM Regressor, to predict market volatility using the Optiver trading dataset. Evaluated models based on MAE, RMSE, and R² metrics, with Random Forest Regressor showing the best performance due to its ability to handle non-linear data and consistent residuals. Enhanced the Random Forest model through advanced data handling and feature engineering to further improve predictive accuracy and interpretability. Visualized predictions using scatter and residual plots to assess model performance and identify areas for improvement.
Optiver Trading Challenge: Market Volatility Prediction
Utilized various regression models, including Linear Regression, SGD Regressor, Random Forest Regressor, and LGBM Regressor, to predict market volatility using the Optiver trading dataset. Evaluated models based on MAE, RMSE, and R² metrics, with Random Forest Regressor showing the best performance due to its ability to handle non-linear data and consistent residuals. Enhanced the Random Forest model through advanced data handling and feature engineering to further improve predictive accuracy and interpretability. Visualized predictions using scatter and residual plots to assess model performance and identify areas for improvement.
Unet-PLUS | Oral Cancer Image Segmentation
Developed deep learning models for oral cancer image segmentation using U-Net architecture with various pre-trained backbones (ResNet, InceptionNet, EfficientNet) for feature extraction. Achieved high performance using Intersection over Union (IOU) metric and further improved results through a weighted ensemble approach combining the strengths of multiple models. This project demonstrates expertise in image segmentation, deep learning architectures, and medical image analysis.
CONNECT WITH ME
