AI for manufacturing, running inside your company
AI for manufacturing that runs on your own servers: technicians find the answer in the manuals in seconds, and support tickets arrive already classified with a suggested first reply. Manuals, data sheets and ticket history never leave the company.
The knowledge is already there, just buried
In a manufacturing company the knowledge support needs already exists: tens of gigabytes of machine manuals, data sheets and revisions, plus years of closed tickets where someone has already fixed the same fault. The problem is finding it. The senior technician knows where to look, the new hire does not, and whoever routes tickets reads them one by one.
A generic chatbot wired to a cloud service is not the answer. You need a system that knows your machines, answers by citing the manual page it took the information from, and runs on your own hardware, because technical documentation and customer history are company assets.
Four pieces, each useful on its own
Technician assistant over the manuals (RAG)
Manuals, data sheets and service notes (typically 10–100 GB) are indexed in a semantic search engine. Technicians ask the way they would ask a colleague and get an answer that references the document and page.
A dataset from closed tickets
Ticket history becomes a clean dataset of problem → solution pairs: deduplicated, anonymised where needed, with normalised categories. It is the least visible piece and the one that decides the quality of everything else.
Ticket classification with fine-tuning
An open-source model with 8–14 billion parameters is trained on the dataset to assign category, machine and priority to each new ticket and draft a first reply in your support team's style.
Integration with your ticketing system
Classification and draft reply show up directly in the ticketing system you already use, via API or connector. The operator confirms, edits or discards: a person keeps the final say.
From the first use case to the integrated system
Free intro call (30 minutes)
You tell me which machines, how many documents, how many tickets and which ticketing system you use. I tell you whether the project makes sense and where I would start.
Prototype on manual search
We start with a subset of manuals and a group of technicians who actually use it. If the answers are not useful, the project ends here with a contained investment.
Dataset and fine-tuning for classification
We build the dataset from closed tickets, train the model and measure it on tickets it has never seen, against manual classification.
Integration and production
Ticketing integration, permissions, logs, periodic index updates when new manuals come out and retraining as the history grows.
Orders of magnitude
50,000 tickets come to roughly 25–50 million tokens of text. Fine-tuning an 8–14 billion parameter model with techniques such as LoRA on a dataset that size takes a few hours on a professional GPU, not weeks. Hardware is sized on the actual case: for daily use (manual search and classification) a server with one or two GPUs is usually enough, not a data centre.
RAG for documents, fine-tuning for style and classification
| RAG (document search) | Fine-tuning | |
|---|---|---|
| What it is for | Making the model answer from the manuals | Teaching a format, a style, a classification |
| Source data | Manuals, data sheets, service notes | Closed tickets: problem → solution, category |
| When new documents arrive | Update the index, that's it | No retraining needed |
| Cites the source | Yes, document and page | No |
| Typical mistake | Indexing duplicate or outdated documents | Using it to make the model "learn" the manuals: it makes things up |
Private AI: your data never leaves the company
The whole system (model, document index, ticket dataset) runs on company servers or on dedicated infrastructure you control. No document is sent to an external service for processing, and no ticket ends up training someone else's model. The models are open source, so there are no per-token costs that grow with usage and no dependence on a vendor who can change prices or terms.
About me
I'm Stefano Salvucci, a registered engineer (Order of Engineers of Rome) with 12+ years across software development, architecture and automation, and co-founder of BusinessinCloud. I work as a fractional CTO: I decide with you what is worth building, then I build it. Based in Rome, I work with companies across Italy and abroad; site visits happen in person, the rest remotely.
Common questions about AI in manufacturing and support
How is AI used in manufacturing, in practice?
The fastest returns come where there is a lot of text to read: searching machine manuals for technicians, classifying support tickets, drafting replies to customers. These projects work on documents the company already has, with no sensors to install or production lines to change.
What does on-premise AI mean?
The AI model runs on company servers (or on dedicated infrastructure under the company's control) instead of a third-party cloud service. Manuals, data sheets and tickets stay inside the company perimeter.
Is an on-premise LLM as good as ChatGPT?
For focused tasks, yes: answering from your manuals with RAG and classifying your tickets after fine-tuning are jobs where an 8–14 billion parameter open-source model does well. It is not meant to write poetry or answer on any topic, and it does not need to.
Why not train the model directly on the manuals?
Because fine-tuning teaches style and format; it does not reliably memorise content, and a model fine-tuned on manuals answers confidently even when wrong. To make the model know the documents you use RAG, which retrieves the right passage and cites it. Fine-tuning is kept for classification and reply format.
How many tickets do you need for fine-tuning?
It depends on how many categories there are and how clean they are, but a history of tens of thousands of closed tickets (50,000 tickets is about 25–50 million tokens) is more than enough material. The real work is often cleaning the dataset, not the training, which takes a few hours.
Is this the same as AI predictive maintenance?
No. Predictive maintenance works on machine sensor data to forecast failures. This project works on documents and tickets to help the people who fix those failures. They are complementary, and the second is usually faster to put into production.
Does it integrate with our ticketing system?
Yes, if the system exposes an API or supports integrations, as most ticketing and help desk software does. Classification and draft reply appear where the operator already works, with no extra app to open.
How do we start?
With a free 30-minute intro call. You tell me about your machines, documents, ticket volumes and tools, and I tell you whether and where I would start.
Let's talk about your manuals and tickets
30 minutes, free: we assess whether and how AI can help in your case.