The Clash of Clans War Strategy and Resource (COCWSR) Optimization Model leverages neural networks to optimize resource allocation and war strategies. This model detects base layouts, troop compositions, and attack patterns to offer real-time recommendations for efficient upgrades and tactical decisions. By providing detailed analysis and predictive insights, it empowers players to make data-driven strategies, enhancing both gameplay and overall efficiency.
EcoHarvest is a web application designed to enhance sustainable farming practices by harnessing advanced data analysis techniques. It uses scikit-learn's tree regression model to optimize crop selection, revitalize soil nutrients, and improve farmer profits. Additionally, EcoHarvest integrates APIs such as Google Cloud, OpenAI, Twilio, and Meteostat to deliver real-time weather updates and personalized recommendations, achieving a 96% model prediction accuracy (PENNAPPS XXIV - University of Pennsylvania)
Cipher Grid offers a secure graphical QR code generation and decoding system. By combining graph theory and SHA-256 hashing, it ensures robust encryption and efficient decryption, making it ideal for secure data transmission and authentication. This approach provides enhanced protection for sensitive information, making Cipher Grid a reliable solution for secure authentication and confidential data sharing. (MATH 485 – Graph Theory)
Designed and deployed EDURA, a Moodle-integrated AI chatbot that replaces static AAC quizzes with realistic, multi-turn role play conversations for training pre-service teachers in active listening (LAFF) strategies. Built with GPT-4o-mini, Whisper, and ElevenLabs, EDURA simulates diverse parent personas with authentic emotional tones, evaluates teacher responses against the LAFF rubric, and delivers real time grading with auto-generated PDF feedback. The fully modular plugin integrates seamlessly into Moodle’s gradebook, supports persona customization through an admin UI, and currently supports 5,000+ pre-service teachers across 12+ institutions worldwide, all at under $0.001 per session for highly scalable adoption.
FiboTune explores the intersection of mathematics and music by generating MIDI compositions from numerical sequences such as the Fibonacci sequence or any user-defined formula or list of numbers. The system maps these sequences into musical notes and rhythms, enabling users to hear mathematical patterns as melodies. Beyond simple sequence mapping, FiboTune can also allow users to upload their own music files to guide the generation process, adapting the output toward a chosen style. This approach provides both structured and personalized results, making the compositions unique to each user’s inputs. Users can experiment with mathematical creativity while still producing music that aligns with familiar patterns.
Contributed to NuGraph3, a hierarchical Graph Neural Network for reconstructing neutrino interactions in Liquid Argon Time Projection Chambers (LArTPCs) with the UChicago DSI × Fermilab. Focused on object condensation and scalability using MicroBooNE data, implementing architectural updates by initializing β (clustering confidence) and coordinate embeddings in encoder for richer context-aware learning. Reduced training time from 14 hrs to 2 hrs per epoch (7× faster) via multi-worker asynchronous loading, pruning redundant computations, and migrating to NVIDIA H100 GPUs with mixed-precision CUDA. Analyzed clustering quality trade-offs (ARI, loss) while processing 10TB+ of simulation data, bridging HPC and particle physics.