SunnyData has been our go-to Databricks partner that helped us move this project from pilot to production.
Client Challenge
Manual brand review could not keep pace with the catalog
Arcos Dorados operates McDonald's across 20 countries in Latin America and the Caribbean. Every product image it publishes reaches customers through three channels: the mobile app, in-store digital kiosks and third-party delivery partners. Each new image had to be validated against McDonald's brand standards (framing, lighting, approved cups, valid food and beverage pairings, prohibited text) before going live.
The process was entirely manual. Each image opened a ticket in InvGate and waited for a member of the brand team to review it. Approval took 134 hours on average, was available only during business hours, and 16% of tickets submitted off-hours waited 113 hours for approval. Manual review cost $70k a year, and global campaigns generated demand peaks the team could not absorb. Arcos Dorados needed same-day approval with no human in the loop, and a way to add new markets without launching a new project each time.
The Solution
AI-powered image moderation on Databricks
SunnyData designed and implemented an end-to-end moderation pipeline on Databricks that detects new tickets, validates each image against market-specific brand rules, and closes the loop automatically. The MVP went live in Argentina and was rolled out to 15 markets in only 5 months.
Databricks Workflows poll InvGate every hour of every day, and pick up new tickets with the attached image and market metadata.
Eight validation modules run in parallel against the rules of that market and the active campaign, using Claude served through Mosaic AI Model Serving.
Approved images are published to the production bucket used by the catalog app. Rejected images return with a written reason. The ticket closes automatically.
The solution runs entirely on Databricks, deployed as code. Adding a market means editing one configuration file with its rules, cups and beverages; the model and pipeline code stay unchanged.
The Agent at Work
Eight validation rules, applied consistently to every image
Each module evaluates one dimension of McDonald's brand standards. Rejected images are returned with a written explanation, so teams can correct and resubmit the same day.
Key Benefits Achieved
Documented business results, before and after
Approval times are measured from InvGate ticket timestamps, and the $70k saving was validated by the business. The production image bucket is also governed for the first time: 73% fewer images in Argentina and 74% in Chile, organised automatically by channel and asset type.
A replicable accelerator for distributed catalogs
The pipeline is packaged as a reusable accelerator: infrastructure as code, a market configuration file for the rules and a governance table for exceptions. The same pattern applies to any brand with a distributed product catalog and centrally defined standards.
Tell us where you are. We'll show you what's possible.