Aushadh Grid

Each medicine line is simulated day by day over its own resupply window (7 to 14 days) against its forecast. A transfer arrives the day after dispatch at the earliest; the donor gives up the stock today. Demand that cannot be met on a day is counted as lost, because patients go without rather than wait. Warehouse indents arrive with routine resupply, so they are not counted here.

Moved in calm times, before any alert, so the stock is on the shelf when the surge starts. The weakest district and scenario are served first. A donor gives only what it would not need in its own worst emergency, so no donor becomes less prepared. Moves stay within 150 road km, and cold-chain items only move between working fridges.

Today each PHC holds its own safety stock against its own bad days. When the district's PHCs share one virtual shelf, a surge at one PHC is covered by quiet shelves nearby, delivered the same day by the district vehicle from the hub. The same availability then needs far less safety stock, because bad days rarely hit every PHC at once (square-root law, 95% service). Where even a shared shelf falls short of 90% in some emergency, the hub holds a buffer that it issues oldest-first into routine use, so the buffer rotates instead of expiring.

Detection uses the CDC EARS C2 method: today's count is compared with the mean and spread of a 7-day baseline that ends 2 days earlier, so a slow rise cannot hide itself. Three standard deviations above baseline on 2 of the last 3 days raises an alert. An alert can switch the demand forecast into outbreak mode for exactly the districts affected, days before an outbreak is officially declared.

Forecast: Holt linear smoothing on day-of-week adjusted consumption (120 days). Before fitting, days when the PHC had run out are replaced with estimated true demand (a stock-out records zero use, which would teach the model that demand is low), and data-entry spikes are capped. Where a PHC's report is overdue, today's stock is estimated from forecast use since the last report and the uncertainty is widened.

Forecast

Select a row to see its forecast.

A lot that a quiet PHC cannot use before its expiry date is sent to a busier PHC that will use it in time. That PHC then draws on this stock first, and its own later-expiring stock lasts longer.

Two kinds of transfer. Prevent stock-out moves surplus to PHCs that would run out before resupply. Vital medicines are served first, nearest donor first, and donors with expiring stock are preferred. Rescue expiring stock moves lots that would expire unused to PHCs that will use them in time. Cold-chain items (2–8 °C) move only between PHCs with a working ice-lined refrigerator, within 150 km.

Each district vehicle leaves its headquarters, collects from donor PHCs, delivers to receivers and returns. A transfer joins a run only if that costs no more road km than a separate trip, and a run stays within an 8-hour day (35 km/h plus 15 minutes a stop). Runs carrying vaccines, insulin or oxytocin need a cold box.

One SMS per PHC, in the same medicine codes pharmacists already use to report stock, so it works on any phone. PHCs whose last report is 3 or more days old are asked to count their stock before anything moves, so nobody drives stock around on stale numbers.

PHC capacity register

CSV columns: phc_id,medicine_code,stock, with optional expiry_days (days until the lot expires). Medicine codes: . Rows replace the stock on hand and mark the PHC as reported today.

Each national node trains the demand model on its own PHC data. Only model weights leave the country; patient and facility records stay local. The coordinator averages weights by sample count (FedAvg) and sends the global model back. Nodes with little data, and little outbreak history, gain the most.