Built AI system that turns handwritten roofing orders into priced quotes with zero manual steps

Achievements

Built for a paying client — Marko of Virumaa Katused OÜ, an Estonian roofing and sheet-metal company. A photo of a handwritten sheet-metal order slip goes in; a priced quote, a cutting plan, a bending drawing and an open-questions sheet come out as PDFs with no manual steps between (proof run 8 Aug 2026, audited). The owner approved the first real output set on 6 Aug 2026. The pricing formula was machine-validated against 1,228 articles of the client's accounting data; a replayed historical case priced to the cent against the real invoice (€78.54). In a blind test on 7 real slips, all 16 cut-widths comparable to invoices were read correctly; zero wrong numbers passed as confident — uncertainty became questions instead. The pattern is already repeating for a second client.

Tools

Claude as vision reader, driven by a versioned prompt bound to live Airtable rules; Python for everything deterministic — geometry, standard cut-width selection, pricing, PDF generation; Airtable as the single home of 59 production rules, the price catalog and the owner's review interface; 63 synthetic tests plus machine gates checking every AI reading before pricing and every shape before rendering. There is deliberately no n8n: the bottleneck was domain logic, not workflow plumbing. n8n-class tooling enters the next phase: a self-service quoting portal for small contractors.

Methodology

One fact, one home: every operational rule lives in a single Airtable table; scripts read generated snapshots, and a hardcoded business constant is treated as a bug. Source material is immutable; test answers stay sealed until after each run. The human review gate is a design decision, not a gap: the acceptance criterion is "no wrong number may pass" — the system stops and asks where uncertain; the owner reviews a control sheet before anything goes out.

Learnings

A full audit (8 Aug, 47 findings) caught the chain drifting from its own doctrine — VAT and production constants hardcoded in ~13 places, PDF branch locked to one slip type; fixed the same day. A worse failure was mine: during the blind test, seven parallel readers shared one temp folder; two readings nearly cross-contaminated, the readers detected and reported it themselves. Both slips have since been re-read cleanly — results confirmed. Airtable's MCP cannot create forms — the intake form stays unbuilt, e-mail intake comes first; automation boundaries are worth mapping early. Baseline: ~300 quote requests by early August this year; a profile quote takes the owner 30–120 minutes by hand, target 10 — not yet measured, the pilot measures it. Full-roof quotes (1–4 hours) are the next phase.

Relevant URLs

None public — a private client system. Artifacts, the audit report and the Airtable base can be shown on request.