← Back Case studies

Three builds, written up properly.

One is a production system a client still runs. One is an n8n template with no client attached, included for what it took to survive three unreliable APIs. The third is my own tool, with a live install it checks against itself. Every number is as it ran. Client names are out wherever there's a client to protect.

01 · Client system · live in production

Five order types, one weekly job

A subscription meal kitchen in Japan, roughly 65 subscribers and about 600 meals a week. Orders arrived five different ways and somebody spent four to five hours every Monday turning them into a cook list and a shipping list by hand.

The system reads around 5,000 records, produces the prep list, the shipping summary and a print-ready PDF per box, and holds back anything it can't verify rather than guessing. Some of those orders carry allergy information, which is why guessing was never an option.

~30 minWas 4–5 hours

12,500+Meals processed

24+Weeks running

0Weeks I've touched it

Read the full case study →

Loading walkthrough

02 · n8n build · submitted as a template

A video pipeline that survives three unreliable APIs

Takes an Instagram reel, finds every cut, pulls a frame from each scene and rebuilds the whole thing shot for shot around a different product. Same shot count, same pacing.

No client on this one, it's a template I built and submitted to n8n's Creator Hub. It's here because of what it took to make it reliable: every slow step submits a job and polls rather than sitting there waiting, and three of the APIs behaved differently from their own documentation.

10,324 msRebuilt timing, exact

28Nodes

3API contracts corrected

Read how it was built →

03 · Own build · open source, live demo online

The failures are quiet. This finds them first.

An installer and verification toolkit for an agency running the same AI receptionist install for many small businesses. One command sets a client up, a second runs thirteen automated pass/fail checks against the live install.

No client on this one either, but there's a live install to check: a fake plumbing company this tool provisioned and verifies against itself. It's here because building the checks turned up more real bugs than writing the product did.

13Automated pass/fail checks

9Work against any install, not just this one

3Real bugs only running it ever caught

Read the full case study →

Interactive · runs in your browser

Run the compiler yourself

Rather than describe the review queue, here it is. I rebuilt the kitchen system's compile logic against invented orders. Hit run, see what it refuses to process, fix those rows and watch the reports rebuild. No client data is anywhere near it.

Try it →

Your process won't look like either of those.

That's fine. The judgement about what to automate and what to hold back for a person is the same either way. Twenty minutes and I'll tell you whether it's worth doing.

Book a call Email me instead

You'll be talking
to me, not a team