Python, FastAPI, Rasterio (GDAL), WebSockets, GeoPandas, Mapbox GL JS
- Built an event-driven intelligence system streaming Sentinel-1 SAR sub-scenes directly from AWS Earth Search via GDAL Virtual Rasters using dynamic STAC queries without local disk caching.
- Implemented a two-pass CFAR anomaly detector with spatial Non-Maximum Suppression (NMS) and global vector land-masking to isolate vessel radar cross-sections from ocean clutter and coastlines.
- Fused radar detections with live WebSocket AIS broadcasts rendered targets at 60 FPS in Mapbox GL JS using decoupled GeoJSON sources.
Production Full-Stack Application | Next.js, React, TypeScript, Google Gemini API, Supabase (PostgreSQL)
- Developed a live web platform that evaluates receipts to determine whether an expense qualifies as a legitimate business deduction and maps it to the correct Canadian tax category.
- Built a multimodal AI pipeline to extract line items, calculate provincial sales taxes, and generate contextual audit-defense justifications explaining why each expense qualifies.
- Architected a secure cloud backend in Supabase/PostgreSQL with Row Level Security, passwordless authentication.
Algorithmic & Natural Language Processing System | Python, spaCy, Graph Algorithms, Spaced Repetition (FSRS)
- Created an educational language tool that evaluates free-form student writing by grammatical syntax rather than rigid, word-for-word templates.
- Applied dependency-tree parsing and recursive tree traversal to isolate grammar mistakes from vocabulary substitutions.
- Integrated a spaced-repetition algorithm that tracks grammar micro-skills to dynamically serve exercises targeting specific conceptual weaknesses.
Data Engineering & Database System | PostgreSQL, PostGIS, Node.js, Python, Puppeteer
- Built an automated ETL pipeline that scraped, parsed, and standardized messy offshore well log files across legacy energy-sector repositories.
- Designed an extensible relational database schema in PostgreSQL/PostGIS to store arbitrary downhole sensor curves without breaking database structure.
- Evaluated string-similarity and N-Gram algorithms to match non-standard sensor abbreviations against an industry dictionary with 90.7% accuracy.