The system that runs Octus Forge

Octus Forge is my company, and this is the system it runs on. An order comes in and moves through seven stages until it’s a sealed box — intake, modeling, approval, production, painting, packing and shipping — and every stage has an owner and a checklist. Around that flow sit finance, inventory, the product catalog, and an audit trail of who changed what. The front end is React + Vite on Vercel; behind it is AWS: Cognito for identity, Lambda, and Postgres on RDS with row-level security. It started on Supabase and was migrated. STL upload takes files up to 5 GB, because 3D models of large parts get that big.

The most useful fix in the project didn’t come from an automated test; it came from the first employee. When she sat down to run a real order, it was clear within minutes that she couldn’t — and shouldn’t — push the flow to the end on her own: decisions I’d left open were decisions only the order owner can make. An internal approval gate went in before the critical stages.

The same session exposed a worse defect. When a write failed, the screen said it had worked and the data vanished on the next reload — the person redid the work without knowing why, and on the system side nothing was logged as an error. Silent failure in software people use all day erodes trust faster than any visible bug. Now the operation rolls back and says so.

The system is in production, with employees running real orders every day. The repository is the company’s and private.