Case Study
From Hackday to USPTO Patent
A machine learning photo auto-tagging system built at Salesforce — years before AI went mainstream.
USPTO
Patent Granted
ML
Vision API
2019
Filed
0
Manual Tags
The Challenge
Photo metadata is broken.
Professional photographers shoot thousands of images per session. Every one of those images needs metadata — keywords, categories, scene descriptions, subject tags — to be searchable and sellable. Without it, photos disappear into archives and never get licensed.
The standard workflow is manual: a photographer sits down after a shoot and tags each image by hand. It's tedious, inconsistent, and the first thing that gets skipped when deadlines hit.
Stock agencies reject submissions with poor metadata. Editorial teams can't find assets. Entire shoots lose value because nobody tagged them properly.
The question was straightforward: could machine learning do this automatically — and do it well enough that photographers would actually trust it?
The Origin
A hackday at Salesforce.
This started as a hackday project at Salesforce. The idea: connect Adobe Lightroom — the industry-standard photo management tool — to Salesforce Einstein Vision, their machine learning platform for image classification.
Instead of manually tagging photos, the plugin would analyze each image through the Einstein Vision API and automatically generate keywords, categories, and scene descriptions. A photographer imports their images, and the metadata writes itself.
The prototype worked well enough that Salesforce's patent team took notice. What started as a one-day experiment became a full patent application. Luke and co-inventor Paulson McIntyre filed the patent together.
How It Works
Four steps. Zero manual tagging.
01
Import
Photographer imports images into Lightroom as usual. The plugin hooks into the import workflow automatically.
02
Analyze
Images are sent to Salesforce Einstein Vision API for classification — scene detection, object recognition, and category prediction.
03
Tag
ML-generated labels are mapped to IPTC/XMP metadata fields — keywords, categories, scene descriptions — and written back to Lightroom.
04
Refine
Photographers review and adjust tags as needed. The system learns from corrections to improve future predictions.
The Innovation
ML before the AI boom.
This was built years before ChatGPT, before DALL-E, before "AI" became a boardroom buzzword. Machine learning image classification existed, but nobody had connected it to the professional photography workflow in a practical way.
The patent covers the full system: the integration between desktop photo management software and cloud-based ML APIs, the metadata mapping layer that translates ML predictions into industry-standard IPTC/XMP fields, and the feedback loop that improves accuracy over time.
Patent Details
Patent Number
US20210241019A1
Title
Machine Learning Photographic Metadata
Co-Inventors
Luke MacNeil & Paulson McIntyre
Assignee
Salesforce, Inc.
The Results
From experiment to patent.
USPTO patent granted for ML-powered photographic metadata (co-invented)
Automated tagging eliminated hours of manual metadata work per shoot
Demonstrated production ML integration years before the current AI wave
Bridged the gap between enterprise ML platform and creative industry workflow
Keep Reading
Related articles.
Blog
Building AI Infrastructure for Production
How to build AI systems that actually run in production, not just demos.
Case Study
Autonomous AI Podcast
Another ML pipeline — fully autonomous content production with zero manual steps.
About the Engineer
Luke MacNeil
20+ years at Salesforce & CVS Health. USPTO patent holder. Published technical reviewer.
Ready to build AI that ships?
From hackday prototypes to production ML systems — let's talk about what AI can do for your business.
Book a Technical Assessment