When COVID hit the UAE, customers started ordering from their local bakkala (the neighbourhood grocery) by WhatsApp voice note instead of walking in. When restrictions lifted, the behaviour stayed — voice notes had quietly become the default way to order, and the volume never went back down.

Each order arrives as free-form speech — "send me two cases of water and a box of tea" — often in Emirati Arabic, English, or a mix of the two, and spread across several voice notes and texts. Turning that into a clean order by hand is slow and error-prone: listen, interpret, match every item to the right catalogue reference, get the quantities right, all day.

Worse, the words rarely pin down a product. "Send me milk" — full-fat or low-fat? Which brand, which size? The right answer usually isn't in the message at all; it's in the shop's standing order guide (its cadencier), not in any single order — and sometimes the exact item is out of stock and needs a substitute.

And it has to hold up under real conditions: messages arriving faster than they can be processed, race conditions between messages in the same conversation, out-of-order delivery, and a different catalogue and live inventory for every merchant.

We built a voice-to-order pipeline. Whisper transcribes the audio (transcriptions run in parallel per conversation), in Arabic or English, with English↔Arabic translation so a note in either language resolves to the same products; a classifier then extracts the items and quantities, and RAG matches each one against the merchant's own catalogue.

When a request is ambiguous, we resolve it the way a good shop assistant would: rank the likely products from the merchant's standing order guide (its cadencier), ask a short disambiguating question when confidence is low, and suggest a sensible substitute when the exact item is out of stock — so the order still goes through.

The hard part is the ordering, not the transcription. A Redis task queue processes each conversation's messages in strict sequence to avoid race conditions, aggregates consecutive messages into a single order, and keeps matching consistent with the merchant's live inventory — so several voice notes become one correct order instead of several conflicting ones.

The system runs on GKE (Kubernetes), built in Go with pooled gRPC connections between services; only the GPU transcription runs on Cloud Run, scaling to zero between bursts to keep it affordable. Structured error handling and Sentry keep it observable in production.