Projects
A tool for flooring contractors that automatically scans large architectural PDF packages and extracts floor plans, elevations, finish schedules, and sheet indexes. Aiming to automate the estimating process to maximize efficiency and accuracy.
Runs a three-tier classification pipeline per page: embedded text extraction, Tesseract OCR fallback for scanned drawings, and Claude Vision as a last resort for unresolved pages. Handles the ~50/50 split between clean CAD exports and messy scans.
Focusing on AI-powered feedback tools that give students accurate, granular commentary on their assignments — going beyond a grade to surface exactly what was strong, what fell short, and why. Aligned to instructor grading standards to make automated feedback feel as consistent and intentional as a human reviewer.
Achieved 5% better alignment with instructor rubrics via structured JSON output schemas enforced across 150+ assignment examples. Built a synthetic data pipeline to stress-test edge cases and validate model behavior.
End-to-end sentiment analysis for movie reviews — surfacing exactly what worked and what didn't for movie studios. Goes beyond positive/negative to deliver aspect-level insights that make review data actually useful.
Currently building this into a customer-facing full-stack application for any domain.
SDSU students needed a better way to handle car accidents in tight campus parking lots. The Crash App enables easy, incentivized incident reporting — including hit-and-runs — to build a safer campus community.
Built at Innovate 4 SDSU Hackathon 2025.