*You don’t learn to use AI by reading about it. You learn it the way you learned your ERP: by doing real work with it, getting it wrong and trying again.*
The training day ends at four. The slides were good, the demo was impressive, and everyone nodded at the right moments.
On Monday morning the planner opens the forecast file. The AI tool is right there. And nothing happens, because the demo used someone else’s data, someone else’s question and someone else’s company.
That gap between seeing and doing is the reason this site works the way it does. Everything here comes with something to try, and it all happens at the same company: Upshift, a bicycle maker that doesn’t exist.
What this site is
This is a personal knowledge-sharing project. I’ve spent more than twenty years in supply chain and logistics, and the last few years using AI in that work: first for questions, then for tools, then for agents. The site is where I write it down.
There are three kinds of things here:
- Blog posts on new models and developments, and what they mean for supply chain work.
- Books, starting with a core series on using AI, building tools and creating agents, plus specialist books that go deeper into one decision.
- Tools and guides, such as the model directory, the project planner and the guides to building your first app.
The rule behind all of it is simple. Nothing goes up without something you can do with it this week: a prompt to run, a calculation to rebuild or an exercise with a way to check your answer.
Why a fictional company
An exercise needs data, and every choice of data has a catch. This is how I weighed it:
| Your own company’s data | A toy example | Upshift | |
|---|---|---|---|
| Safe to share and paste | No | Yes | Yes |
| Realistic mess | Yes | No | Yes, on purpose |
| Consistent from one exercise to the next | Yes | No | Yes |
| Big enough to matter | Yes | No | Yes |
| Everyone works on the same case | No | Yes | Yes |
Your own data is the real test, but you can’t put it in a book, and you shouldn’t paste it into a tool you haven’t checked. A toy example is safe, but it teaches the wrong lesson: real supply chain data is never ten tidy rows.
Upshift sits in between. It’s a complete company with its own history, so a lesson learned in one place still holds in the next.
Meet Upshift
Upshift designs, paints and assembles road and gravel bikes, most of them as a regular and an e-bike version. It has two plants, in Eindhoven and Bydgoszcz, and five distribution hubs, from Venlo to Basel. It buys from 30 suppliers, ships with 8 carriers and serves 52 dealers, retail chains, distributors and web shops across Europe.
Behind it sits one synthetic dataset of 49 connected tables: products, orders, stock, production, transport and more. The dates follow a date you choose, so the company always looks current. Some of the mess is deliberate, because a clean dataset would teach you nothing about the checks you need.
You can read the full introduction, and see how the data fits together, on the Upshift page.
How one question grows
The same question can be worked at three depths, and that’s how the site is built.
- In a blog post, it’s a prompt. In the Sonnet 5.5 post, a warehouse review had ten rows of data with two errors hidden in it. You can paste it into a chat and see in ten minutes what a model catches.
- In an exercise, it’s the whole company. The same kind of question runs on the full dataset, with thousands of rows and a known right answer, so you can check your result instead of trusting it.
- In a book, it’s the method. The chapter explains why the checks matter, what to hand to the model and what to keep, and how to make it part of your week.
You can start at any depth. Most people start with a post, and go deeper when a question sticks.
What a fictional company can’t do
- It isn’t your data. Your mess is different from Upshift’s mess. The exercise builds the habit; your own work is where it proves itself.
- Synthetic data is too tidy in places. I add realistic faults on purpose, but real data will always find new ones.
- It reflects my view of supply chain. A bike maker with plants and hubs covers a lot, but not everything. Retail, pharma or process industry would look different.
- It won’t make the decision for you. The exercises train judgement, not replace it. Checking the model’s assumptions is part of every one of them.
Tell me what to build next
This site grows with the questions people bring to it. I’d like to hear yours:
- Exercises you want to see. A process at Upshift you’d like to work on, or a task from your own week you’d like to practise.
- Topics to focus on. A model, a tool or a technique you’d like explained for supply chain work.
- Where you got stuck. An exercise that was unclear, too easy or too hard.
- What helped. It tells me what to do more of.
Use the Feedback button at the bottom right of every page and choose Idea. You don’t need an account, and every message comes straight to me.
Learning AI is no longer about watching someone else use it. It’s about doing real work with it, somewhere safe enough to make mistakes. That’s what Upshift is for.
So tell me: which exercise would you like to see at Upshift next?
