[ DELIVERY_LOG ] · --:--:-- UTC · edge · …
Four production systems.
Four industries. One operator.
Every project below is live on its own domain, doing real work for someone who is not me. This page is the record of how they were built: what the model actually did, what stayed human, and where it got things wrong the first time.
[ METHOD ]
The part that decides whether it works
is not the code generation
Anyone can get a model to produce a working page. The difference between a demo and something a business runs on is whether the domain rules inside it are correct — and whether anybody checked. Six stages. Select one.
Stage 02 and stage 05 are where the value is. Skip either one and you have produced a very confident spreadsheet that is wrong in a way nobody notices for three months.
[ CASE_FILES ]
Four builds,
nothing in common but the method
Enterprise storage sizing. Civil-works estimating. A sports-coaching programme. A florist's shop. Different clients, different stakes, the same line running underneath.
[ LIVE_TOOL ]
Scope one of your own processes
this is the workshop, compressed
Not a chatbot. This is the first hour of a real engagement: you describe a repetitive process, and you get back a structured automation spec — every step with an owner, the model tier it should run on, the guardrails, the points where a human must sign, the KPIs to watch, and a deliberately conservative estimate.
The spec appears here. A new process takes ten to fifteen seconds — it is a 70 B model generating against a schema, not a lookup table. The three examples above are already in the edge cache and come back instantly.
Runs on Llama 3.3 70B through Workers AI, in the Cloudflare data centre nearest you — no origin server involved. A fresh process takes ten to fifteen seconds; identical prompts are then answered from the edge cache in under a tenth of a second and cost nothing, which is the cheapest FinOps lever there is and the reason the examples feel instant. Treat the result as a first draft of a scoping conversation, not a plan.
[ RUNTIME ]
How this page itself runs
the same rules apply to my own work
There is no origin server behind this site and nothing to keep patched. One Worker serves the static assets and the single API route, from whichever of Cloudflare's data centres is closest to you.
- —data centre serving you right now
- —measured round trip to the API
- 1Worker — assets and API in one deployment
- 0npm packages at runtime, nothing to CVE-patch
- 0build steps — the source is what ships
This is not minimalism for its own sake. A dependency tree you do not maintain is a liability with a delay fuse, and every build step is a thing that can break at three in the morning between you and a fix. The four projects above run the same way.
[ OPEN_CHANNEL ]
Got a process that
everybody retypes by hand?
That is usually the first line worth building. Send me the awkward one — the process that is too small for a software project and too annoying to keep doing.