Case study
A complete bespoke business system for a Hungarian trading-and-manufacturing company — from specification to production in 12 days.
A complete, bespoke business system for a Hungarian trading-and-manufacturing company: from AI interpretation of the incoming order e-mail through approval, documents, stock and delivery, to matching incoming payments. It runs in production, without a single interruption.
Measured: August 2026.
Every module is in live use. The list comes from the user manual, which documents only behaviour that actually shipped.
The central workflow: order interpreted from e-mail, approval, documents, delivery, settlement. A three-panel view — the list, the incoming letter and the order data side by side.
Quotes with their own numbering, versioning and expiry tracking — as a phase of the same case, not a separate register.
A speaking article-number system by size, colour and finish, with colour variants, packaging units, partner-specific prices and stock levels.
Stocktaking and stock levels, usable on mobile — in the warehouse, not at a desk.
Terms, individual prices, credit limits with warnings, payment deadlines, agent commission — per partner.
Generating, sending and tracking proforma invoices, delivery notes, work sheets and invoices — every issued document in one place, with status.
Automatic matching of incoming transfers to orders and documents, with operator approval and splitting across several documents.
Dispatch planning and a driver view on mobile: own runs, handover confirmation, movement between sites.
Ask the system questions in natural language, run operations by written instruction — permission-bound and audited.
Screens from the working system, on fabricated data. The company names, people, addresses and amounts are all invented — no real customer or partner data appears in them.
01
Nine days went into processing meetings, correspondence and domain walkthroughs before a line of code was written. Every specification started from the structured knowledge base, never from raw notes — and that is what made four days of development possible at all.
02
The system was built with multi-track, parallel AI orchestration: every work stream in its own worktree behind its own quality gates — build, test, end-to-end — and only work that passed got merged. The 811 commits are throughput, not haste.
03
Go-live happened on day 12, and from then on the system has worked on real orders. We measure the accuracy of the e-mail interpretation layer by replaying live orders against what the operator actually approved — not on a synthetic sample.
This project produced the interpretation layer we are now developing as a standalone product — and for which we are looking for pilot partners.
Article numbers from e-mail — the layer, on its own →