AI systems, ML infrastructure, and applied research.

Built a college-marketplace product across mobile, backend, infrastructure, payments, trust, and admin workflows.

• Architected Dockerized Node.js microservices on RDS and Nginx with JWT auth, HTTPS, health checks, and operational separation.
• Built Android/iOS search and media workflows with PostgreSQL full-text search, filters, S3 signed uploads, compression, and Rekognition.
• Customized Vue admin tooling with RBAC, analytics, ticket transfers, proof galleries, and zoomable receipt review.
• Launched real-time chat, Stripe payments, bidding auto-expiration, and QR/6-digit pickup verification with photo evidence.

Worked on infrastructure observability, runtime visualization, OCR triage, and development pipeline isolation.

• Visualized infrastructure runtime across Kubernetes clusters and reduced idle PostgreSQL database connections by 70 percent.
• Built OCR-based text detection and triage from input video using OpenCV, then improved runtime with multiprocessing.
• Implemented a development pipeline that separated testing workflows from production runs.
• Integrated visualizations for message-queue wait times using external APIs and custom internal scripts.

Built software, firmware, data processing, and ML analysis workflows for EEG-based cognitive and affective-state experiments.

• Developed an EEG software interface with Stroop, math, and video-feedback tests in Python.
• Designed and tested firmware for a custom EEG collection ADC board in C++.
• Conducted experiments with mentors, collected data through custom EEG software, and processed results in Jupyter Notebook.
• Used NumPy, Pandas, and SciPy for data merging, trimming, and feature extraction.
• Applied SVM, KNN, and random-forest classifiers to EEG data analysis for emotion and cognitive-state detection.

Asynchronous LLM Reinforcement Learning Under Constrained Hardware

A from-scratch AReaL-style async RL system for Qwen 2.5-0.5B on 2 V100 GPUs, studying bounded staleness, policy lag, and learning efficiency under constrained hardware.

OmniPaint Reproducibility Study for Object-Oriented Diffusion Editing

An ICCV 2025 reproducibility study of OmniPaint, validating removal metrics on the public benchmark and exposing insertion benchmark sensitivity with public substitute data.

Robust Perception for Autonomous Vehicles: Camera-Only vs. Camera+LiDAR Fusion

A CARLA closed-loop robustness study comparing YOLOv8n camera-only perception with RGB+LiDAR PointPainting-style fusion under weather, viewpoint, and adversarial stress.