A photo becomes a defensible number
Between a shopper's ten-second snapshot and a ministry's enforcement decision sits a small stack of models: reading tags, notarising receipts, resolving items, rejecting fabrications, measuring what controls actually did — each one obeying the aggregate doctrine.
Snap to signal, safely
Shelf tag in, observation out
Vision models read the price label through the plastic strip: item, pack size, price, promotion state — normalised to the national item taxonomy in under two seconds, on a phone in a market with one bar of signal.
The receipt is the notarisation
A tag photo says the shelf claimed a price; the receipt says it was paid. Line-item extraction cross-checks both, timestamps the transaction and upgrades the sighting's confidence tier — the difference between a complaint and evidence.
Same chicken, every shop
Item resolution across brands, weights, grades and local names — 'ayam bulat', 'ayam kampung', family pack, 800 g truncation — collapsed onto one comparable series so the weekly movement is real, not taxonomy noise.
Gaming the grid
Fabricated sightings have signatures: photo reuse, impossible geotags, burst patterns, temperature-inconsistent timing. Anomaly models quarantine before aggregation and bend a region's published confidence score down — loudly, on the dashboard.
What the basket tells the ministry
Elasticity and substitution reads on top of the index: where a price control moved shelves or only moved the queue, which items are spilling into which, where an inspection trip is statistically worth making.
Density where officers aren't
Coverage models predict which districts are about to go thin this week and nudge contributor incentives there — the same dynamic-sensing pattern that keeps a hazard feed honest, aimed at baskets instead of rain.
How far the models are not allowed to go
Point, never convict
Models surface bands, movements and anomalies for ministry humans to act on. AnHarga never files a case, never names a store to enforcement — the index's authority depends on that distance.
Aggregate before intelligent
Every model output that leaves the vault is cohort-floored and differential-private first. A clever read on exposed individuals is not a feature; it is the incident the PDPA amendment was written to punish.
Confidence is published
Every region, item and week carries a visible confidence score. An index that hides its own uncertainty is propaganda — this one ships its error bars next to the numbers.
Every model here obeys a floor test written years before the first tender. See where the doctrine came from.
Why this shape — the evidence floor→Ready to build something intelligent?
Let's engineer your next product — from idea to launch.