experiments.sh
██████╗ ██████╗ ██████╗ ██╗███████╗ ██████╗████████╗███████╗ ██╔══██╗██╔══██╗██╔═══██╗ ██║██╔════╝██╔════╝╚══██╔══╝██╔════╝ ██████╔╝██████╔╝██║ ██║ ██║█████╗ ██║ ██║ ███████╗ ██╔═══╝ ██╔══██╗██║ ██║██ ██║██╔══╝ ██║ ██║ ╚════██║ ██║ ██║ ██║╚██████╔╝╚█████╔╝███████╗╚██████╗ ██║ ███████║ ╚═╝ ╚═╝ ╚═╝ ╚═════╝ ╚════╝ ╚══════╝ ╚═════╝ ╚═╝ ╚══════╝
user@lab:~/experiments$ whoami
Shengguang Cui // AI Engineer & Researcher
user@lab:~/experiments$ ls -la | wc -l
5 experiments found
user@lab:~/experiments$
> Displaying 5 of 5 experiments found
experiment_01.log
🔒
Jan 2025 - Mar 2025[COMPLETED]
$ Automated Privacy Testing for LLM through Fuzzing
// Course Project - Trustworthy AI, UCLA
>Built a privacy harness to test LLMs' PII extraction from HTML profiles
>Extended PROMPTFUZZ with privacy-oriented mutators and HTML-aware templating
>Uncovered a +7pp lift in attack success rate (85% → 92%) on GPT-4o
PythonLLM SecurityFuzzingPrivacy Testing
experiment_02.log
🧠
Feb 2023 - Dec 2023[COMPLETED]
$ HeteroPruneFL - Federated Learning Framework
// Research Project - Duke Kunshan University
>Designed heterogeneity-aware FL framework with client-specific subnetworks
>Introduced dynamic sparse training (prune–regrow) for local data adaptation
>Achieved consistent accuracy gains across 4 datasets vs. 3 baselines
PyTorchFederated LearningNeural Network PruningDistributed Systems
experiment_03.log
🖼️
Mar 2023 - Apr 2023[COMPLETED]
$ Campus Image Generative Models
// Course Project - Advanced ML, CUHK
>Collected and curated CUHK-Shenzhen campus image dataset
>Trained GAN, DCGAN, and diffusion model with ImageNet pretraining
>Confirmed diffusion outperformed GAN in realism and mode coverage
PyTorchGANDiffusion ModelsComputer Vision
experiment_04.log
📹
Jun 2025 - Jan 2026[IN_PRODUCTION]
$ Live Captioning & Translation System
// Industry Project - Videospace, Inc.
>Reduced caption latency from 6s to 1s for Fortune 500 clients
>Implemented real-time translation and AI highlights generation
>Deployed Whisper and PaddleOCR models with GPU auto-scaling
PythonJavaWowzaSpeech-to-TextOpenAI API
experiment_05.log
👁️
Mar 2025 - May 2025[COMPLETED]
$ Hallucination Reduction in LVLMs
// Research Project - UCLA
>Proposed contrastive decoding approach for factual content steering
>Implemented and benchmarked on LLaVA-1.5 with POPE dataset
>Demonstrated hallucination reduction vs. baseline methods
PyTorchLLaVAVision-Language ModelsContrastive Decoding
contact.sh
🤝
$ cat ./collaborate.txt
Open to Collaboration
Research opportunities, open-source contributions, or interesting AI/ML projects. Let's build something together.
$mail -s "Collaboration" lilil@ucla.eduPID: 5642UPTIME: 984dMEM: 83%LOAD: 0.91