ESREAL Replication & Extensions
Cyclic image-to-text pipeline with InstructBLIP, SDXL, GroundingDINO, and token-level rewards for hallucination mitigation.
Filter and explore. Click a card for details.
Cyclic image-to-text pipeline with InstructBLIP, SDXL, GroundingDINO, and token-level rewards for hallucination mitigation.
Compose UI, multi-touch first-player picker, animated selection, clean game HUD.
Orientation game where players identify campus locations from photos, GeoGuessr-style.
Classic Minesweeper with server-based sessions, difficulty levels, and a top-10 leaderboard.
Extracts subreddit posts and comment threads using Yellowcake, and shows clean JSON in a simple web UI.
PostgreSQL-backed order processing system for call intake and manufacturing workflows.
Benchmarking multilingual LMs on bitext mining and retrieval-based classification across 200+ languages.
End-to-end loop: image -> caption (InstructBLIP) -> reconstructed image (SDXL/SD3.1) -> region alignment checks (GroundingDINO) -> token-level rewards. Focus: reducing hallucinations and improving factual grounding.
Player picker using multi-touch input with animated selection, stateful UI, and clear layout tokens.
Orientation game where players guess campus locations from photos. Built as a course group project with a focus on navigation, UX, and production-readiness.
Players connect via IP, port, and username, choose a difficulty, and play a full Minesweeper session with bombs, flags, and a top-10 scoreboard.
Extracts subreddit posts and comment threads with controls for max posts, comments per post, and sorting. Results display instantly in the browser with copy-to-clipboard.
Built a PostgreSQL-backed order system used by call intake and manufacturing teams. Implemented order, customer, and status workflows plus search/reporting to improve handoffs.
Benchmark spanning 11 datasets and 200+ languages, evaluating bitext retrieval and retrieval-based classification. Shows competitive results via semantic retrieval without fine-tuning.