DOSTI field ops

A US-registered nonprofit whose Pakistan office has no reliable internet: field staff placing, collecting and auditing donation boxes across Peshawar and Khyber, and five years of school-aid handovers kept in one Excel sheet.

Data dying in transit. Field entries with nowhere to go when the signal dropped, and quantity cells Excel had quietly turned into negative numbers.

1. 01

   Dosti Impact writes to IndexedDB first and syncs deltas with a cursor when the network comes back. Photos queue in their own store with orphan pruning and client-side compression; duplicate collections are guarded.

2. 02

   WhatsApp receipts retry until they land, and moved off the AiSensy aggregator to Meta's Cloud API directly.

3. 03

   The DOORS importer is idempotent: it parses Excel's mangled quantity cells, fills a missing district only when every clue agrees, and reloads a corrected sheet without duplicates.

4. 04

   Data migrations run behind a write guard with a mandatory dry run. The outreach dashboard reports only what the forms actually record.

1,130

donation boxes tracked

production export, 20 May 2026

281

tests: unit, integration, IndexedDB persistence

1,245

deliveries imported

Nov 2020 – Apr 2026

10,161

units recovered from 311 mangled cells

96%

district inference accuracy

in backtest

PKR 15 → 2.9

per WhatsApp receipt

per-message rate, Sept 2026

• An earlier outreach headline was withdrawn as inaccurate once the importer showed what the forms record: about 7,500 students across 1,245 deliveries. The smaller number is the true one.
• I have not measured a sync-reliability rate, so I do not quote one.
• District inference is 96% in backtest, not 100%. That is why the importer leaves a district blank unless every clue agrees.
• PKR 15 → 2.9 is a per-message rate, not a measured total saving, and Meta reprices Pakistan from 1 Oct 2026.
• The code and data belong to DOSTI, so there is no public repository.