PocketDokani

Small online sellers in Bangladesh run their shops from chat, handwritten labels and spreadsheets. An order is negotiated in Messenger, then the recipient, final price and COD are retyped into a spreadsheet and courier form, and whoever packs it has to ask again. PocketDokani is a phone web app that covers this flow. It turns conversation screenshots and voice notes into source-linked order drafts, checks a parcel photo against the confirmed order, prepares the courier booking, and answers business questions in Bangla, Banglish or English. AI prepares the work; the seller approves it. Each drafted value links to its source quote, code compares labels against the order, a failed parcel check blocks booking, numbers are rendered from stored records, and changes need explicit confirmation. Development run on one real chat and the seller's real workbook: extraction 8/8 fields on 3/3 repeats, business questions 20/20 correct with 0 confidently wrong, wrong labels caught 7/7, wrong bags caught 3/3, clean parcels passed 5/6 (one unsure). Limits: these sources were used during development, so this is not held-out accuracy. Parcel labels were constructed; no real handwritten label, real parcel photo or live shipment was tested, and no time saving was measured. A photo cannot verify sealed contents. Built during the hackathon. The seller's spreadsheet already existed and was used as a private data source. Judges can try an isolated sample shop without logging in.