DukaanVoice PK — ڈیجیٹل دکان — is an offline-first point-of-sale and shop register for Pakistan’s ~3 million kiryana (neighbourhood grocery) shops, with an AI accountant called Munshi AI at its core.

Small shops run on thin margins and unreliable internet. DukaanVoice turns the counter into a fast, auditable register and gives the owner a Munshi that understands Urdu, Roman Urdu and English (for example, “2 kilo atta bill mein daalo”), answers with tool calls over the shop ledger, forecasts stockouts and recommends reorders, flags cash-desk mistakes and unusual expenses, quotes live wholesale mandi rates, and speaks answers aloud.

The register includes Quick Sale, Dashboard, Products, Transactions, Udhaar Khata (customer credit), Expenses, Cash Closing (Roznamcha) and Reports. Money is stored as integer paise; confirmed sales decrement stock exactly once; cash closing excludes digital payments; and a network outage does not stop a sale.

The AI principle: the model proposes, deterministic code disposes. Open-weight Gemma parses Urdu free text into confirmable draft bills, while deterministic code performs the arithmetic. TabPFN forecasts 7-day product demand; SerpApi provides live mandi rates; ElevenLabs speaks answers; Sentry traces tool calls; and Mastra orchestrates tools. The app degrades gracefully offline, and its 87-test suite runs without keys or network.

Open models matter here because local inference can keep customer names and credit balances on the shop’s device, avoid per-query costs that are difficult to absorb on grocery margins, and better handle local vocabulary such as Roman Urdu, “ser” and “pao” units, and local brands.

Watch Video
link text

v2
v3