Machikoba no Techo July 2026

What is connected is not a cable. It is sight.

July. Summer has arrived in earnest here in Arita, and the heat builds inside the factory from early morning. On rainy days the humidity adds to it, and the air turns heavy.
In northern Kyushu the swordtip squid season is at its peak, and the night sea is lit with the lamps of the fishing boats. Beside that sea, our factory goes on making egi as it always has. Cut, fit, cut again, adjust. Every body is washed one at a time before it is assembled. Twenty-seven years since we were founded, and it is this plain, unglamorous handwork, repeated, that becomes each single egi. We are sincerely grateful for your continued support.

Meanwhile, outside the workshop, hardly a day passes without hearing the word "AI".
For about a year now I have been running my own experiments with a range of AI tools.

But the change I feel is not that AI has become clever.
It is that a small factory can now build, in-house, the kind of system it previously had no choice but to order from a specialist firm.

Let me describe one example.

■ A factory is full of displays

A factory has displays everywhere.
The control panel of a machine tool. Thermometers. Hygrometers. CO₂ sensors. Power meters.

People look at them and judge:
"It's running a little hot today."
"That reading is off compared with usual."
"About time we ventilated."

Getting that information into a computer used to be hard work.
Request the communication specification from the manufacturer. Run a communication cable. Install a PLC. Buy dedicated software.
In other words, you first had to connect to the machine physically.

Machining centres and NC milling machines do publish API specifications, and with enough determination you can pull the data out in C++. But the labour involved is enormous, and it is hard to summon that determination.

■ Don't touch the machine. Just look at it.

Recently, though, another route has opened up.
Film the display with a webcam.

Not a special industrial camera. The one we used is an ordinary webcam that cost about 3,000 yen.

Read the numbers off that video, store them as data, plot them, and raise an alert when something is wrong.
It converts what a person used to take in with their eyes into data that can be analysed continuously.
Put another way, we applied DX to the act of observation itself.

■ The first build, and the wall it hit

The first version was very simple: send the captured image to an AI and ask it to "read these numbers".
Anthropic's Claude. Google's Gemini. xAI's Grok. We tried a number of services.

On top of that, coding assistants such as Claude Code, Codex, Grok Build and Cursor have made building the system itself dramatically faster. Work that a specialist would once have spent weeks on now runs in a short time.

But once the prototype was running, a problem surfaced. It cost too much.

A single call is not expensive at all. Monitoring around the clock, all year, is another matter. "AI you use occasionally" and "AI you keep running" demand completely different cost structures.

As the owner of a small factory, more than being impressed by the performance, I find myself thinking about how to make the monthly invoice smaller.

■ Running it locally

So the next attempt was to run a local LLM on a Raspberry Pi.
A large PC next to a machine tool is in the way; a Raspberry Pi that fits in your palm sits there comfortably.

And if the AI runs entirely on our own equipment, there are no API fees. The running cost is effectively the electricity. In theory, close to zero.

Reality was not so simple. Most of the well-known LLMs would not run — not enough memory. In the end Qwen ran most comfortably, though it still would not work through convenient tooling such as ollama and needed a fair amount of tuning.

■ Use as little AI as possible

Where we finally arrived was this:
OCR handles the normal case. The local LLM starts only when OCR fails.

Lightweight OCR does the everyday work. AI appears only when something is abnormal or the reading is ambiguous.
The running cost came out at essentially zero.

To get speed out of a Raspberry Pi, the standard approach is to fix the camera position and hard-code the coordinates to be read. That approach has a weakness. A slight vibration in the factory shifts the camera, the hard-coded coordinates no longer line up, and suddenly nothing can be read at all.

Here again the context-aware LLM earned its place. Because it can re-read "where the numbers are" from the whole image, a small shift is easy to correct. We built that correction into the reading routine and had cron run a periodic calibration. The result is a system that keeps absorbing drift on its own, with nobody watching over it.

The interesting part: a project that began by trying to use a lot of AI became most practical once it evolved towards using as little AI as possible.

It resembles how people work. Routine jobs are handled by experience and procedure. You call in a specialist only when something unexpected happens. That division of labour turned out to be the most efficient one.

■ Not by cable, but by sight

Thinking this through, I found myself remembering how we charge a phone.
It used to be obvious that you plugged in a cable. Now wireless charging is everywhere — you set the phone on the pad and it charges. There is no physical connection. Energy still crosses.

This project is a little like that.
Traditionally you attached a communication cable to get information out of a machine. Not this time. We do not touch the machine. There is no data line. We only observe. And still we get the information we need.

You could call it a shift from a physical cable connection to a contactless, induction-like way of picking up information.

We did not connect to the machine. We observed it.
What is connected is not a cable. It is sight.

Come to think of it, sight works through light, and light is also an electromagnetic wave.
Seen that way, perhaps this really is a connection by electromagnetic induction, like a transformer.

■ The cost of trying has fallen

The biggest thing about this project was not that AI got clever. It was that the cost of attempting something fell dramatically.

A system we would have commissioned from a specialist a few years ago can now be built inside a small factory. We do not have the budget of a large company. We have no dedicated department. But precisely because we are small, we can try, fail and improve quickly.

I believe that is where the real change of the AI era lies.
Not that AI takes work away from people. That AI has greatly widened the range of what people can attempt.

KEYSTONE will go on taking up the problems that arise on the factory floor — in egi making above all, but not only there — adopting new technology flexibly as we go.

Traditional yet innovative.
With those words in mind, we intend to keep moving forward, one step at a time.

Thank you, as always, for your support.

with SQUID!!

July 2026
KeyStone Corporation
Nobuhide Kanagawa, Representative Director