Blog/Case notes
A photo of the shelf is worth a morning's walk
Notes from the shelf-intelligence system running at a supermarket chain in Thailand. The hard part wasn't the vision model.
1 min read · · Boring Agents
- 1Phone photos beat new cameras: no hardware, no installation, and staff already carry the phones.
- 2The hard part is the catalogue. Matching what's on the shelf to 32,000 products is most of the work.
- 3A list of gaps is noise. A list ranked by the sales each gap is costing is a morning's work.
The usual pitch for shelf monitoring is cameras on every aisle. That means hardware, installation, a maintenance contract, and a conversation with whoever owns the ceiling. We went the other way. Store staff photograph the shelves with a phone, the way they already photograph things for head office.
The photos go through a pipeline overnight. It finds each product on the shelf, works out what it is, and compares what it sees with what should be there. That middle step is the real job. The chain's catalogue has more than 32,000 products, many of them in near-identical packaging. Telling two flavours of the same crackers apart from a hurried photo under fluorescent light is where most of the effort went.
Then we stopped thinking like engineers. A list of every gap in a store is long and nobody reads it. So each gap is matched against the store's sales, and the morning worklist is ranked by what the gap is costing. The manager starts at the top and stops when the morning runs out.
None of this needed a breakthrough. It needed the catalogue to be right, the photos to be good enough, and the output to fit into a store manager's first hour. That is most of applied AI.
Tell us the job your team hates. We'll tell you if AI can do it.
A 30-minute call with the people who would build it. No technical knowledge needed: just describe the job, how often it happens, and how you check it's done right.