Shengguang Cui

GPA: 4.0/4.0

Courses: Information Theory, Computational Robotics, Deep Learning, Trustworthy AI, AI on Chips, Neural Signal Processing

GPA: 3.7/4.0 · Rank: 5/100 in major

Courses: Computer Architecture, Operating System, Parallel Computing, Database Systems, Algorithm Analysis, Machine Learning

• ●Shipped a live captioning system for Fortune 500 clients which reduced caption latency from 6s to 1s by engineering a hybrid Java-Python pipeline on Wowza Streaming Engine to pipe resampled audio to speech-to-text services and convert output to ID3 tags for player rendering
• ●Enhanced live webcasts with real-time translation and live AI highlights by developing a Python generation module utilizing OpenAI and Gemini APIs; added dynamic user controls for language and vocabulary by routing configuration updates through Wowza's REST API endpoints
• ●Scaled production inference by deploying in-house Whisper, PaddleOCR, and LLaMA models on Lightning AI with GPU auto-scaling and API serving capabilities, while evaluating LoRA and RAG strategies to implement continuous knowledge injection in production
• ●Built a Slack integration module with Django models, views, and REST API endpoints to manage OAuth authorization and asynchronous message retrieval, enabling users to search and view LLM-generated summaries of relevant discussions from Slack workspaces
• ●Architected an asynchronous Message Processing Engine using Celery to execute tasks dispatched from Django API endpoints; engineered a distributed pipeline to fetch Slack messages, segment conversations into topic clusters, and generate structured summaries and titles

• ●Researched hallucination reduction in Large Vision-Language Models (LVLMs), focusing on inference-time methods
• ●Proposed a contrastive decoding approach (based on On-the-Fly Preference Alignment) to steer LVLM outputs toward factual content
• ●Re-implemented Activation Steering Decoding (ASD) as a baseline and developed a PyTorch implementation of the proposed method; benchmarked on LLaVA-1.5 with POPE and LLaVA-Bench, demonstrating the proposed method reduced hallucinations relative to ASD

• ●Designed HeteroPruneFL, a heterogeneity-aware Federated Learning framework that assigns each edge device a client-specific subnetwork via importance-based pruning, enabling training within restricted compute/memory budgets while preserving comparable model performance
• ●Introduced dynamic sparse training (prune–regrow) in client-side training to adapt per-client network topology to local data within memory caps
• ●Engineered a reproducible PyTorch stack (server–client orchestration, autorun scripts, analysis pipeline) and benchmarked on 4 datasets vs. 3 baselines, showing consistent accuracy gains under different resource constraints

• ●Built a privacy harness in Python to test LLMs' PII extraction from HTML profiles; extended PROMPTFUZZ with privacy-oriented mutators (context injection, persona/role prompts, obfuscation, pseudocode) and HTML-aware templating to generate adversarial prompts automatically
• ●Uncovered a +7 pp lift in attack success rate (85%→92%) on GPT-4o through iterative mutation and response analysis

• ●Collected and curated a CUHK-Shenzhen campus image dataset; applied preprocessing and augmentation to build a high-quality training set
• ●Trained GAN, DCGAN, and a diffusion model (with ImageNet pretraining) in PyTorch; conducted a comparative study using direct and t-SNE/LLE visualizations, confirming diffusion outperformed GAN and DCGAN in realism and mode coverage