Twelve times the volume in five weeks, with quality holding at 95 to 98%
Human verification of food classification across thousands of near identical items for an AI food waste platform.
Illustrative image, not client data
The challenge
Client D uses computer vision to track food waste for grocers and prepared food operations. Its models had to tell apart thousands of items, many nearly identical: overripe produce, dishes in similar packaging, food at different stages of spoilage.
More than 50 stores generated over 10,000 image logs a day, all needing human verification within store hours across US time zones.
What we did
We ran a two week pilot with 2 to 3 annotators to prove feasibility and calibrate quality, then scaled to 12 annotators within five weeks.
Annotators worked inside Client D’s own platform, correcting misidentified items across a catalogue of thousands.
Results
Daily volume grew from 4,000 to more than 12,000 logs, and quality held between 95 and 98% throughout.
The engagement is on hold after Client D changed its service model.
