Fabbrica AI by Moneta · the intelligent factory
It learns your data and builds you its model.
A small model for each machine, retrained every night on its logs. At the edge, no GPU. Zero data scientists, zero prompts.
Your machine's model
< 1 M parameters · industrial PC · retrained every night
The animated stories
The intelligent factory. Every machine learns, the factory reasons.
A local LLM beside every PLC, trained every night on its logs. Titanium, the central brain, asks them all.
How it works · illustrative example
The small brains talk to the big one.
- 01
The PLCs write
a log line every instant
- 02
One LLM per machine
retrained every night, no GPU
- 03
Up to Titanium
state, anomalies, shift summary
- 04
You ask
Titanium asks the 15 machines
The production manager
"Why did Line 2 lose 6% yield this week?"
2 machines out of 15 have a signal
Flow-pack 3 · right jaw −4 °C since Tuesday: 140 packs scrapped
Labeller · offset +1.8 mm since Thursday: 60 manual reworks
The other 13: no signal.
Titanium answers
Two causes: Flow-pack 3's right jaw, cold since Tuesday (70%), and the labeller's offset (25%). Preheat the jaw and recalibrate the offset.
Who sees the drift, and when · Flow-pack 3, right jaw, three weeks
Ordinary ML
day 18
One sensor, one threshold. It sees the fault when it is already there.
Generative AI
never
It knows everything about the world, nothing about this jaw.
Local LLM
day 12
It knows the machine's normal, it sees the drift. And it names the cause.
Before and after · every format change
From two hours of trials to ten minutes.
2 h
10 min
Changeover time
300
20
Scrap per changeover
6
1
Trials on the machine
Before, the technician tries on the machine, by feel. After, the AI compiles the recipe upstream and the PLC runs it.
Why the AI stays out of the loop
The PLC in milliseconds, an LLM in seconds.
A delay is a phase shift: the loop oscillates. So the AI compiles beforehand, and the PLC stays inside the cycle.
The architecture
The AI prepares the recipe, the PLC runs it.
In the example a new format arrives, the BISC150 biscuit. The AI works beforehand, calmly, away from the machine. On the machine the PLC stays, fast and predictable as today.
1 · Before, in the office · hours or days
The history
the 40 recipes already made, scrap, logs
The AI proposes
three recipes for the new format
Tests them on the computer
1 million simulated cycles
The technician
picks, approves or corrects
2 · Then, on the machine · milliseconds, every cycle the same
Loads the recipe
for the format in production
Checks
that the values are within limits
Runs
moves chain, film and jaws
The training
Last layers or prompt? Each where it errs the least.
"The prompt describes the machine in words. The layers learn it from its numbers."
How much it makes up, alone
Orders of magnitude · by distance from the history
Inside the history
At the edge
Outside
Outside the history it sends you to the physical trial. Inside, verification stays.
1Proposes
3 candidates
2Simulates
1 M cycles · 99.1%
3The PLC checks
170 °C · 130 ppm · 70 ms
4The technician decides
approves, corrects, signs
We improve predictive AI
A generic ML model or a general AI lose the machine's details. Its own model does not.
We know the machine from its numbers: the last layers learn that line's wear, settings and materials. Four published results say why it works.
Daily training
Every night it restarts from the day's logs.
The machine changes: wear, new materials, settings. A model trained once stays stuck at how it was. Its own model retrains every night, at the edge, and follows it.
Mon
retrains
Tue
retrains
Wed
retrains
Thu
retrains
Fri
retrains
Sat
retrains
Sun
retrains
Day's logs → last layers retrained → model updated by morning. No GPU, nobody has to restart it.
Prediction error over time
illustrative example+10 points
Small and trained beats the generalist
310 models with 7 billion parameters, fine-tuned on task data, beat GPT-4 by 10 points on average across 31 tasks.
LoRA Land, technical report 2024
0
LLM advantage on time series
Removing the LLM from time-series forecasters does not lower accuracy: pretrained models do no better than models trained from scratch.
NeurIPS 2024 · Tan et al.
+30%
The model that knows the machine
On field data, the model adapted to the real machine gains up to 30% accuracy over generic machine learning under noisy conditions.
Applied Sciences 2025
↑
New facts in the prompt, patterns in the layers
Training an LLM on facts it does not know raises hallucinations linearly: that is why datasheets and manuals go in the context, the machine's numbers in the last layers.
EMNLP 2024 · Gekhman et al.
Let's talk
What machine is it, and what logs does it produce?
Zero data scientists, zero prompts. In a free 30-minute call we tell you what its model can learn and how fast.