The enterprise GIS, digital twin pipelines, and custom software behind one of the largest refineries in North America. The drone program is where the data starts. Most of the job now is what happens to it after it lands.
I take the refinery's incoming data from CAD drawings, survey, drone captures, and LiDAR, put it in the right coordinate system, and publish it as standards-compliant enterprise GIS in ePlot and as reality capture in the Cintoo digital twin. The site's spatial data has one source of truth instead of five that disagree.
Each airframe earns its place by reaching somewhere the others can't: 400 ft above a flare stack, sealed inside a storage tank, or docked on a rooftop running missions unattended.
The heavy-lift workhorse. Long endurance and swappable payloads for site-wide mapping, structure inspection, and the missions that need the biggest sensors in the air the longest.
A compact quad carrying wide, telephoto, and radiometric thermal cameras in one gimbal. The everyday aircraft for thermal inspections and quick-turn captures. The 4TD is weatherized to live in the Dock.
A weatherproof drone-in-a-box that houses the M4TD and launches it on a schedule. Recurring inspection routes fly unattended, and the aircraft returns to charge itself between missions.
A ducted FPV quad with full prop guards for close-quarters interiors and structures. First-person control and 360° capture where a full-size aircraft won't fit.
The aircraft are delivery vehicles. The value is in the sensors and in reading what they return.
Garyville is a live refinery processing over half a million barrels a day. Every flight happens over active process units, around flammable product, and inside airspace and ground hazards that leave no room for improvisation.
An automated Python pipeline that takes drone LiDAR out of Site Scan and lands it in the Cintoo digital twin. It reprojects from State Plane feet to metres with the correct vertical datum, merges thousands of tiles into a single compressed file, and cleans the result. I also run Cintoo data management for the site, including format conversions for outside engineering firms, and built a 360° camera to point cloud workflow that feeds the same twin.
Detects AprilTag markers in raw drone imagery and georeferences them directly from the photo's camera pose, skipping the photogrammetry pipeline entirely. Getting it accurate meant finding an undocumented quirk in how DJI reports gimbal orientation. Fixing it took position error from roughly 96 ft down to 2 to 3 ft. Packaged as a desktop tool for the team.
Started as a dashboard for refinery leadership showing what the drone program covers and produces. It grew into a builder: upload a plot plan and it generates a full dashboard with a 3D viewer for that site.
Authoritative monument layer cross-referencing State Plane, NAD83(2011) geodetic, NAVD88 and ellipsoidal heights, and plant grid. Every other dataset on site gets validated against it.
Generates KML mapping grids for every Marathon refinery, so capture planning starts from a consistent, correctly placed layout.
Drag a raw 360 video onto one file and walk away. About an hour later a navigable 3D walkthrough opens on its own.
Splits large geotagged captures into processing-sized batches by what the camera was looking at, so oblique photos stay with the ground they cover. Originals are opened read-only.
Maps ~96 process-unit polygons with sub-unit photo routing, a thermal viewer verified within 0.1°F of DJI's commercial reference tool, and single-file HTML snapshot publishing for vendor handoff. The system of record for how inspection flights get reviewed at the site.
One locked map of live aircraft, METAR weather, and animated NEXRAD radar around the refinery. The Rust backend classifies aircraft behavior, detects airspace conflicts by dead reckoning, and predicts whether an aircraft will be audible on site. Ships as a single 6 MB executable.
When DJI's official SDK failed across camera firmwares, I reverse-engineered DJI's proprietary APP3 thermal storage and FLIR's FFF binary container, re-encoding radiometric data via Planck-equation inversion to under 0.01°C round-trip. Ships as one offline .exe used daily by the inspection team.
Compares monthly drone captures of pipe-support spring cans to detect lever movement, corrosion, and physical damage, then generates color-coded PDF reports. What used to be a manual, photo-by-photo review now runs as a pipeline.