Extending the GABRIL framework (Gaze-Based Regularization) to web agents by engineering a custom Video Vision Transformer (ViViT). Currently training on high-dimensional Atari datasets to prove that gaze-augmented attention mechanisms improve sample efficiency in complex tasks.
Engineered Python pipelines for critical portfolio risk metric calculations. Optimized complex SQL queries on 4M+ record datasets, reducing execution latency by 15% for daily executive reporting. Modernized legacy infrastructure by migrating embedded queries to a version-controlled Git workflow.
Architected a real-time IoT ingestion pipeline for 10,000+ e-bikes (80GB/day). Replaced WebSockets with MQTT and Kafka to handle high-concurrency streams. Implemented stateful processing with Apache Flink and reduced bandwidth usage by 60% using Protocol Buffers.
Built an edge-to-cloud surveillance system in a 36-hour TreeHacks sprint, deploying a real-time YOLO pipeline on NVIDIA Jetson and enriching frames with local VLMs (NanoLLM). Indexed observations into a serverless RAG backend on Modal with Elasticsearch vectors to support ranked, natural-language video search.
Orchestrated a secure, self-hosted multiplayer platform. Integrated custom game assets into a modified open-source frontend, resolving dependency conflicts to ensure a seamless UI experience. Secured public access via Cloudflare Zero Trust tunnels (no open ports).
Reproduced the AlphaZero architecture from scratch. Engineered a dual-headed Convolutional Neural Network (CNN) in PyTorch to predict policy and value, integrated with Monte Carlo Tree Search (MCTS) for lookahead.
Architected a cloud-native content platform using Go and PostgreSQL. Orchestrated containerized microservices on AWS ECS and managed the full CI/CD lifecycle to support production traffic for 50+ active users.
Engineered a polyglot data pipeline to predict stock volatility. Built a high-concurrency web scraper in Go (Colly) to harvest news sentiment, fused with market data via yfinance, and trained a TensorFlow LSTM model to detect market hype trends.