Printing Press — Order intake to quote in under 90 seconds

The Problem

Quotes took the shop between forty minutes and two days, depending on who was at the desk and whether the customer remembered the GSM. Every order arrived differently — a WhatsApp photo of last year's job, a phone call asking for "the same as before but gold foil this time", an email with a half-spec. The bottleneck wasn't pricing; it was translating an irregular request into the four variables the press actually runs on: paper, size, finish, quantity.

The System

A WhatsApp Business webhook and an email inbox feed a parser that extracts a structured spec — stock, GSM, dimensions, finish (lamination, foil, spot UV), quantity — from photos, voice notes, and free text. The spec hits a Postgres pricing table that the owner maintains by hand: cost per sheet by GSM, finish multipliers, minimum run rules, margin bands by customer tier. A quote draft lands in an operator console with the parsed spec, the chosen price line, and a one-tap WhatsApp reply.

The decision that mattered most: the operator stays in the loop on every send, but the system never asks them to type a number. Pricing is the owner's territory and changes weekly with paper rates; the model never invents a price, it only retrieves and explains one. The operator's job shrinks to confirm-or-correct, not author-from-scratch.

What I built

• WhatsApp + email intake with photo and voice-note ingestion.
• Spec parser on Claude with a strict 4-field schema, no free text.
• Postgres pricing table the owner edits in a plain admin grid.
• Operator console with parsed spec, price line, and one-tap send.
• Resend fallback for customers who started by email.
• Audit log per quote: input, parsed spec, price line, sent reply.

Outcome

The production manager went from drafting six or seven quotes a day to confirming seventy-plus. Customers who used to wait until evening got a price before the chai cooled. The owner kept full control of the pricing table — she edits it Tuesday mornings when the paper market shifts — and the system never touches a number she hasn't approved.

What I'd do differently

I underweighted the photo-only intake path in week one. I assumed customers would describe the job in text and attach a photo as backup; in practice eight out of ten messages were a photo plus three words ("same as this, 500 pieces"). The parser worked, but the first version asked too many clarifying questions because I hadn't trained it on the actual message shapes. If I started over I'd ship with a hundred real WhatsApp threads from the shop's archive on day three, before writing a line of the pricing logic. The intake is the whole product; the pricing engine is the easy half.