Portfolio

Scrapp - Waste Classifier

Won Fourth Place in the Congressional App Challenge against ~12,682 teams nationwide
Developed an app to classify waste and eliminate confusion when disposing of items — live at scrapp.app
The user photographs or uploads an image of an item, receives structured disposal guidance (material, disposal route, confidence, and caveats), and finds nearby drop-off locations on an interactive map — including ~690 curated battery drop-off sites across Southern California
Built the backend and AI: a Flask API using GPT-4o-mini vision classification, with a fine-tuned YOLO (You Only Look Once) object detection model available as an additional detector
The frontend is a Next.js app installable as a mobile-first PWA (and can be viewed on screens of any size)

Research conducted with Prof. Hajar Homayouni, San Diego State University

Compared Symmetric vs. Asymmetric Cross-Modal Attention architectures on the VQA v2 benchmark, using frozen & unfrozen CLIP-L/14 and RoBERTa encoders with a trainable ~307M-parameter fusion module
Symmetric architecture reached ~72% validation accuracy vs. ~67% for the asymmetric architecture; the asymmetric model’s training accuracy kept climbing (~71%) while its validation accuracy plateaued, a widening train/val gap pointing to overfitting
Diagnosed a structural gradient imbalance behind the gap and proposed fixes (differential learning rates, extended epochs, warm-starting)
Built and managed the full GCP training pipeline (L4/A100 VMs, GCS data pipeline, Papermill + tmux for headless runs)
Ranked in the top 150 out of 550 teams (top ~27%) on the VQA Challenge leaderboard on EvalAI
Paper submitted to WACV 2027 (IEEE/CVF Winter Conference on Applications of Computer Vision, ICORE A-ranked) — currently under review

Independent Research Project

Classifying Marine Mammals using Convolutional Neural Networks

Authored under the mentorship of Arpit Jasapara of UCLA (a Masters student researcher), and discusses an EfficientNetB5 + ArcFace ensemble as well as several other models. These models were then used to classify individual whales and dolphins in the Happywhale dataset on Kaggle and achieved a rank in the top 11%. An ensemble of models presented in the manuscript ranked in the top 5% out of 5 thousand teams on Kaggle. After undergoing a 3-month peer review proccess, the manuscript has been published in the National High School Journal of Science.

Open Coding Society Scrum Master

Led a team of 6 during the development process of a full-stack web application using Agile Methodologies.

Backend, created using Flask, contained Machine learning models that predicted areas in the city of San Diego that require maintenance.
Frontend, created using Jekyll and Tailwind displayed the outputs of the machine learning models and provided an efficient yet visually pleasing interface for users.
Created a Kanban board to organize key features and tasks during development

Served as deployment admin for a team of 6 during the development of a full-stack social media application for AP Computer Science.

Developed a full-stack application with JWT authentication, features to gain insights on users’ interests, and a deployed backend using AWS.
Managed user account data and information in SQL databases, and deployed a Jinja site that used session tokens for authentication.
Used Agile Methodologies, maintained a Kanban board, a Frontend, and a Backend.

UC Irvine x Gifted and Talented Institute: Game Science Entertainment Technology

Learned about Gaming Computer Science, Game-based learning, and Game Research
Participated in lectures led by UCI professors and industry professionals
Developed a game in a group of 4 and pitched it to potential investors and judges

FlipGun Project and Pitch