The menu was the easy part when I built DineQR. The useful part is the kitchen board. An order that still needs someone to carry it across the room has not saved much work. Guests scan a table QR and order without an app. dineqr.food

MAX_CONCURRENT_JOBS defaults to 1 in my PDF to markdown tool. Docling runs on CPU and a single worker already takes 4 OMP threads, so a second job in parallel just fights the first for the same cores and both land later than if they had queued. Threads per job is the knob worth touching.

The Docker image for my PDF to markdown tool needs two volumes and the second one is easy to miss, /root/.cache/docling. Skip it and every container restart re-downloads 1 to 2 GB of models, so the first conversion crawls and people assume the app is broken. Packaging ML is mostly cache plumbing.

Most apps I build for small businesses have the same shape underneath. A spreadsheet doing a database's job, and someone re-keying numbers between two systems that never talk. Figuring out which system owns the truth usually takes longer than writing the code.