Orto va in città

The customers were there. The system behind them was a phone, a Shopify site and paper notes.

L’orto va in città is a small km0 organic greengrocer in the centre of Naples. Customers were already ordering on WhatsApp, on Simone’s own number, and the shop was drowning in it. Orders were scattered across chats, stock never matched Shopify, delivery cost more than customers paid, and the owner answered the same questions all day. The budget was tiny and the produce is fresh and sold by weight. The brief was not a new channel. It was to keep the conversation customers already liked, and put one reliable system behind it.

A model that can only speak about rows it has just read, and an owner who always has the last word.

The catalogue never enters the prompt. The model calls typed tools that return a few real rows, so it cannot quote a price it has not seen or promise a day the shop does not deliver. Every order waits for Simone’s confirmation before it is final, and 27 of the 104 ready answers hand the chat straight to him. To make this visible we designed Chat di prova: the conversation on one side and, next to it, every call the bot made to the hub. That is how the shop sees whether the bot understood or guessed. The storefront has no checkout either: its cart composes a WhatsApp message, so ordering stays a conversation.

A Medusa.js hub, translated into the shop’s Italian and designed tablet first.

We chose Medusa.js as the hub: open source, no licence fees, native products, stock and orders, and a path to replace Shopify. On top of it we designed the pages a greengrocer actually needs. Cruscotto opens with what needs a decision: orders to confirm, today’s deliveries with a quick «Consegnato», the week at a glance. Magazzino puts problems first, sold out, below threshold, expiring, to sell quickly, then stock and a log that shows who wrote what and from where. Phone and in-store orders go through «Nuovo ordine», which runs the same checks as the bot. Every English string from Medusa was translated or removed.

Customers on WhatsApp, the warehouse on Telegram, and nothing written until someone replies «sì».

The sales bot runs on the shop’s own WhatsApp number, alongside the Business app. It reads the customer, searches the catalogue, proposes delivery days by postcode and prepares the order. The warehouse bot started on WhatsApp and moved to a Telegram group, which keeps the internal channel away from customers and gives groups and buttons for free. Operators write or send voice notes the way they speak: «so’ arrivate 20 kg di fichi». The bot turns that into proposed loads, arrivals and freshness changes, and writes nothing until someone answers «sì». Creating new products from chat is out of scope on purpose.

Rules, answers, zones and costs live in the hub, and change on the next message.

We kept the bot’s behaviour out of the code. Fifteen behaviour rules, a shop dictionary that maps Neapolitan words to products, substitutions, tone and 104 ready answers in ten categories are all edited by the shop in the hub, and take effect on the next message. Delivery follows one rule: postcode first, then days. The bot never promises a day or a time before the zone is known. The 116 served postcodes are grouped into four bands, and the cost lives on the band, not on the postcode, so changing one band moves all its zones at once. Once Simone has confirmed an order, the bot asks how the customer wants to pay.

Plain Italian documents, numbered fix lists, and a beta with real customers.

Before any build we wrote six planning documents in plain Italian for Simone, from the overview to the design spec. The design starts from the Medusa UI kit with the shop’s brand on top: green, Jost for titles, DM Sans for text, soft radii, three breakpoints, and a drawer on mobile because there are too many sections for a tab bar. The build ran in numbered test rounds, each a fix list for our development team where every item carries a code, a link to reproduce it and what to do. In mid-September the beta opened to real customers, announced as experimental, because they ask the questions a team never thinks of. Next: one stock in kilos for weighed products, narrower delivery slots, and closing Shopify.