Starlink consumer platform billing - Angular and .NET full stack development.

- Created a Python framework to easily switch between LLMs and automatically fallback via configs on Firestore.- Wrote multi-threaded Google Cloud Functions to evaluate and optimize the performance of each backend component’s LLM call (up to 42% improvement), using Firestore and BigQuery to read datasets & store metrics.- Made dashboards for each component evaluator using Google Looker Studio, utilizing live metrics from BigQuery.

- Built a state-wide database app for fiber optics management with MySQL as relational database and Streamlit as frontend for users to upload into and view the database.- Prototyped a chatbot using LangChain and RAG to provide internal assistance with domain specific knowledge.

Advised by Professor Wei Wang- Built an end-to-end computer vision pipeline that incorporates contrastive learning and transfer learning with ResNet and Vision Transformer backbone for the USDA using PyTorch.- Trained a deep neural network that achieves 1.472% validation error on yield prediction by leveraging multi-modal data collected through RGB, depth, LiDAR, and multispectral cameras.

Advised by Professor Peter Wonka- Improved a language-vision pipeline for zero-shot 3D semantic segmentation by ~10% mIoU by incorporating state-of-the-art models such as ChatGPT-4, SAM, and DINO.- Explored the augmentation of 2D rendered views with Stable Diffusion (ControlNet) to improve 2D detection/segmentation accuracy before aggregating labeled masks to 3D mesh.- Enhanced image generation dramatically by prompt engineering and writing automated training and evaluation scripts in Bash/Python to perform hyperparameter search (executed via Slurm).

Advised by Professor Stefano Soatto- Used depth estimation as pretraining for vision transformers to improve the performance of visual representation learning applied to downstream tasks like semantic segmentation by 5.7% mIoU.- Configured, trained, hyperparameter searched, and evaluated self-supervised models on the lab’s GPU clusters.- Paper accepted for ECCV 2024 main conference.

- Maintained and implemented new pages for the Regents Scholar website using the React.js framework.- Used HostGator and cPanel to manage the rssla.org domain.

- brainstormed and started building a mobile app using React Native

- Research and adapt machine learning models (CNN, Enformer, etc) for predicting quiescent gene regulation using TensorFlow.- Manipulate and analyze large genomic datasets on the UCLA Hoffman2 cloud computing cluster.

- Learned about then implemented linear and logistical models using NumPy- Using Pytorch to train a CNN that can identify different cassava leaf diseases

- Planned lessons and coding exercises with the professor.- Actively led small group student discussions during class.- Frequently answered student questions both in-class and through email.