idea · Idea
Two thousand handwritten feedback slips a day — drop the folder, get per-store and per-dish ratings. Just over a tenth of a cent per slip.
A restaurant chain gets back a thousand handwritten feedback slips a day — four rating checkboxes and a comment line. In most chains the manager sits down at night and types them into Excel one by one until 11pm, and all that work yields a single total. Which dish actually got the complaints? Nobody has the energy left to count.
Now the manager drags the day's scanned folder into a browser and walks away. Subfolder names become store names automatically. A few minutes later the ratings are already broken down per dish: which store's satisfaction dropped, which dish took 7 negative ratings this week, who wrote a complaint in the comment box — every line opens the original slip image. Reporting to HQ is one export button, with Excel split into a sheet per store.
Running across 14 locations of one restaurant chain for over three months: 60,000+ slips processed, 1,500–2,200 a day. Measured against human review as ground truth: 96.2% of dish-level ratings correct on the first pass, 6 missed genuine complaints out of 9,712 slips (0.06%), and just over a tenth of a cent per slip.
What separates it from "we wired up an OCR API" is one rule: an automatic call has to be right, and anything uncertain goes to a human. Of the slips the machine flags as doubtful, human review confirms the machine was right 95% of the time — and the other 5% is exactly what a person should have looked at. High-risk moves, like flipping a positive rating to a negative one, are never automated. Thermal-paper speckles that a model reads as checkmarks, and upside-down scans that make it read a comment as its opposite, both get caught before the model ever sees them. A complaint is the one thing this system is not allowed to get wrong.
Published on Epoch 0 · by 悠