Foundry OS shipped to a private Turkish foundry in Bursa six months ago. Twelve connected data domains, eighteen functional screens, a seven-stage casting pipeline modelled on the floor's actual vocabulary. This is a field report — not a sales page. What worked, what didn't, what we'd do differently.
What survived week one
Three things landed and stayed landed:
The morning view. A single dashboard showing upcoming work, overdue flags, and a per-company workload breakdown. The owner opens it before his coffee. The order ledger and the events feed sit underneath, so the WhatsApp group quieted within four days.
The 7-stage pipeline. Model Bekleniyor → Model Alındı → Boyama → Kalıp → Döküm → Temizlik → Sevkiyat. We argued internally about collapsing it to five; the foreman flatly refused. He was right. Every stage corresponds to a physical handoff and a quality checkpoint. Hide a stage and the audit log breaks.
The kanban planner. Columns are days. Cards are orders. Drag to reschedule. The floor manager uses it from his phone three times a day. It replaced a paper schedule pinned to the office wall.
What we rebuilt in week two
The quote calculator. We shipped a single-currency MVP. By Wednesday we got a phone call: "the European customer wants EUR with their VAT included, the Saudi customer wants USD without VAT, and our supplier is in TRY." By Friday quotes were multi-currency with per-line tax flags. The schema was right; the UI was wrong.
The defect report PDF. First version was clean A4 with a header logo. The QC engineer asked us to add the operator's name to every row. We hadn't modelled the operator on the report — only the order. That was a one-day refactor we should have done before launch. The lesson: PDFs are part of the product, not an afterthought.
The schema decision we still defend
boyama, kalıp, döküm — the column names are Turkish, the API events are Turkish, the analytics keys are Turkish. We've argued the case for this in the operator-vocabulary essay; the field test confirmed it. New hires onboard faster because the database and the floor speak the same language. Translation is a presentation-layer concern.
We have not regretted this decision once.
What we'd do differently
Ship the photo-evidence upload on day one. We deferred it to month two. Quality reports without photos felt half-finished from the very first week. The work to add it was small; the delay was a mistake.
Pre-build at least one report in PDF and Excel. We shipped PDF only. The accountant wanted Excel for downstream reconciliation. Adding it later cost a week we shouldn't have spent.
Plan the on-call rotation before launch. The first month was bumpy because we were figuring out incident response on the fly. The new rotation runs cleanly — but write the runbook before you need it.
What it cost
Twelve weeks of build, two weeks of pilot, four weeks of stabilization. One full-time engineer plus a half-time designer. Hosted on a single Hetzner box plus a small Vercel project for the marketing site. Total infra bill: under €120/month.
What's next
Multi-tenant rollout to a second foundry in Q3. The schema is templated; the vocabulary stays the same. The hard part is the customer relationship — software is the easier half.
If you operate something specific and the generic ERPs don't fit, send us the brief. We're listening.