The AI is the reason the notebook dies
AnKedai's models exist for one job: delete the typing, the remembering and the month-end panic from micro-merchant bookkeeping — drafted honestly, trained narrowly, confirmed by a human every time.
Photo in. Sentence in. Books keep themselves.
The receipt is the entry
Snap the supplier invoice or the stock pile: vision models post line items, quantities and cost to the ledger in seconds. The notebook dies not because merchants are told to, but because a photo is faster than a page.
Bookkeeping by sentence
“Charli piutang dua puluh, bayar sikit tadi” — voice capture in colloquial Bahasa (and English) resolves into a dated credit entry with the customer attached. The merchant talks the way they already talk to the regulars.
Restock that reads the week
A tiny demand model on the merchant's own history (never anyone else's books) forecasts tomorrow's volume — enough to cut the daily waste that is the kedai's silent third expense line.
AnLebih, priced by the hour
The surplus board needs a price that clears tonight without insulting the morning buyer. A markdown curve learned per stall — the Flashfood/TGTG dynamic window, sized for six bungkus, not six pallets.
E-invoice before the question
Every sale that posts is a valid e-invoice draft: classifications suggested, buyer details completed from history, submission queued to the LHDN gateway through the licensed path. The 2026 deadline becomes an export button.
The books defend themselves
Anomaly detection flags what a human bookkeeper would mutter about: duplicate supplier bills, ghost refunds, staff-till patterns. Alerts are plain sentences, not risk dashboards — the merchant is not a compliance officer.
How far the AI is not allowed to go
Assistant, never auditor
AI drafts; the merchant posts. Nothing auto-sends to a tax authority or a ministry without an explicit human confirm — the honesty spine applies to machines too.
Own books only
Forecast models train on one merchant's ledger, never on the pooled books of peers. Cohort intelligence exists only where the aggregate doctrine allows it.
Small models, honest claims
A kedai's OCR + intent stack runs cheap and offline-first. No frontier-sized claims: the value is the deleted typing, not the parameter count.
The same vision + intent rails already post Masrif’s expenses and i-Saves’ receipts — here, pointed at the kedai.
Why this shape — the evidence floor→Ready to build something intelligent?
Let's engineer your next product — from idea to launch.