The system that runs a defense factory’s production floor

The client is a precision-assembly manufacturer in the defense sector. Parts arrive from suppliers, are assembled at paired workstations and pass multi-stage testing before delivery.

Production, stock, test results and expenses were typed in by hand across spreadsheets, daily progress reports and PDF test reports. Management couldn’t see current stock, how many units could be built right now, or where rejects were happening — and the same output was sometimes recorded twice.

1. [01]
   Stores and inventory: every supplier delivery logged as a batch, oldest stock used first, and a live count of how many units can be built — and which part is holding things up
2. [02]
   Job assignment to workstations, with parts taken all-or-nothing so stock can never go wrong
3. [03]
   Live full-screen displays for every workstation, showing today’s output and the team
4. [04]
   Multi-stage quality testing with pass/fail per stage and defect analysis by reason
5. [05]
   Rejections and rework: every reject has a reason and can be sent back for rework and tracked
6. [06]
   Workforce roster, bench assignments, fingerprint attendance and leave requests
7. [07]
   Expenses and budgets with receipt uploads and budget-vs-actual charts
8. [08]
   Scorecards, production forecasts, batch progress and dispatch tracking
9. [09]
   An AI assistant that writes a graded report every night, checks the numbers every 30 minutes and assigns follow-ups to the right people
10. [10]
    Works offline on web and as a Windows desktop app, then catches up automatically
11. [11]
    17 roles — from assembly worker to COO — each seeing only what they should
12. [12]
    Nightly backups to two off-site locations, and silent auto-updates for the desktop app

1. [01]
   Simple sync over clever sync: the server is the source of truth, and devices queue their work offline and replay it on reconnect.
2. [02]
   Stock accuracy enforced by the database itself — parts are taken all-or-nothing, oldest first — so the numbers can be trusted.
3. [03]
   AI that only runs when something actually changed, uses a cheaper model for routine checks, and can switch to fully local models.

For the technically curious. You don’t need to know any of this to work with me.