A 14-second walk-past video from an ordinary phone was enough for our AI to audit a biscuit aisle against its product list — and find every pricing error and missing product we had deliberately hidden in the data, without a single false alarm.
Anyone can demo AI on easy data. We designed this the way an auditor would: the system was graded against errors it could not have known about.
A single 14-second walk-past of a biscuit aisle in a Dubai supermarket, filmed on a standard phone. No tripods, no lighting, no special equipment.
We built the product reference list from the real shelf — then deliberately broke it: 4 prices changed and 2 products deleted. The list of changes was sealed as the answer key.
The AI compared what it saw on the shelf against the (broken) reference list and reported every discrepancy. Its report was then scored against the sealed key.
The two numbers that matter for a tool store teams would rely on: it caught all 6 planted errors, and it raised zero false alarms — no staff time wasted chasing ghosts.
| Planted error | On the shelf | In the system | AI verdict |
|---|---|---|---|
| Ülker Finger Biscuits 500g | AED 16.25 | AED 14.75 | ✓ WRONG PRICE FLAGGED |
| Ülker Petit Beurre 175g | AED 3.95 | AED 4.95 | ✓ WRONG PRICE FLAGGED |
| Gullon Cracker Cheese 250g | AED 12.95 | AED 14.95 | ✓ WRONG PRICE FLAGGED |
| Nairn's Mixed Berries 200g | AED 14.50 | AED 12.50 | ✓ WRONG PRICE FLAGGED |
| Nairn's Coconut & Chia 200g | on shelf | deleted | ✓ MISSING FROM SYSTEM |
| Julie's Le-mond Lemon 170g | on shelf | deleted | ✓ MISSING FROM SYSTEM |
8 of 9 reference products were automatically located in the footage. The ninth was only half-visible at the edge of frame — and the AI correctly declined to guess rather than report something it wasn't sure of.
Items the AI couldn't verify with confidence weren't forced into an answer — they went to a short review list for a human. Certainty gets reported; uncertainty gets asked about.
Handheld footage, ordinary store lighting, densely packed shelves, near-identical product variants — the same conditions store staff face every day.
This is the actual demo footage with the AI's live detections drawn on top — every product facing and price tag it identifies as the camera walks past.
Shelf-versus-system price mismatches are found in seconds, not discovered by customers at the till — protecting both trust and compliance posture.
The same scan counts facings and spots empty positions, turning walk-pasts into continuous on-shelf availability data.
Manual shelf audits are slow and sampled. A phone pass audits everything in the frame, every time, and writes its own exception report.
Runs on video from the phones staff already carry. No cameras to install, no fixtures to change, nothing to maintain in-store.
This demo deliberately kept scope small: one aisle, a short clip, a hand-built product list. The results justify the next stage — a structured pilot, measured the same honest way.
Multiple aisles of a live store, filmed by store staff during normal walks — no process change beyond pressing record.
Connected to an export of the actual product & price master, so exceptions are measured against the source of truth — store-wide, not a sample.
Success criteria agreed up front and scored with planted-error tests, exactly like this demo — so the results are evidence, not a highlight reel.
We'd welcome the chance to walk your team through the full results and design the pilot around the sections and KPIs that matter most to Choithrams.
Discuss the pilot