Skip to content
Lean-Link

Case studies / Manufacturing

A manufacturer put an AI agent on its own servers, and never sent a byte of production data to anyone.

A manufacturing group with plants in two cities replaced phone calls and manual reports with a local AI agent that watches the line, flags deviations and reports, all on one on-premises server.

0 B

Production data leaving the site, by design

30%

Target operational efficiency gain

20%

Target cost saving

What we built

CUSTOMER PREMISES · BOTH PLANTSSENSORSline · MQTTCOMMS SERVERingestPOSTGRESQLhistoryLOCAL LLM+ automation agentREPORTINGdashboardsALERTSe-mail / SMSTEAMapprovesOpen-source model · no external API · runs on a UPS through a power cut

The distance problem

The problem was never the machines. It was the distance to them. The group runs production sites in two cities, and one operations lead is accountable for all of them. Status travelled by phone call and manual report: by the time a deviation reached the person who could act on it, the shift that produced it was already over.

The company had the data. It did not have it in time, in one place, or in a form anyone could act on.

Why the cloud answer didn’t fit

  • The data would leave the building. Production, yield and quality data is competitive information, and some of their own customers in banking and telecom require that it never leaves the premises.
  • The link isn’t always there. Power cuts and network drops are part of normal operations in the region. Monitoring that stops during an outage isn’t monitoring.
  • The meter never stops. Per-token pricing scales with every sensor reading, forever, on top of import taxes on cloud software licences.
  • Someone else owns the model. A silent version change upstream changes your results, and you can’t audit what you don’t host.

What we built

One server, everything inside the fence. Sensors on the line feed a local comms server, which stores the history and hands it to a local LLM and an automation agent, all running on an open-source model on a single on-premises box, with no external API call. It keeps working through a power cut on a UPS, and it keeps working when the line to the outside world is down.

Agentic AI, kept on a leash

The agent watches every sensor stream continuously, compares it with the plant’s own history and documented tolerances, and judges whether something is noise, a trend, or worth a person’s attention. It raises alerts, opens tickets, and turns a week of raw readings into a page a manager can read in a minute.

Safety-critical decisions stay exactly where they belong: hard-wired into the industrial control system. The agent observes, judges and reports. Anything that would touch a machine is a proposal, and a named person approves it. Every proposal and approval is logged on the company’s own server.

How we ran it

  1. Mapped the current state first. A structured data-collection form, completed with the client before a single line of code.
  2. One site, one question. Not “digitalise the factory”, but a single question with a measurable, valuable answer.
  3. Sized the hardware to the job. One server with a small open model. No GPU farm, no eighteen-month programme.
  4. Human in the loop from day one. Approvals and an audit trail were built into the first version, not bolted on after an incident.
  5. Then replicated. The second plant became a copy of a working thing, not a second project.

Lessons learned

  • Start with the performance question, not the model. AI is an intervention you select, not a place you begin.
  • The environment is a design input, not an excuse. Power cuts, bandwidth and import tax shaped this architecture more than any benchmark did.
  • Small models, well fed, beat big models poorly fed.
  • Decide where the human sits before you build. Retrofitting oversight never works.
  • Confidentiality is something you can sell. Treated as a feature rather than a cost, “your data never leaves” won the deal.

Have a similar use case?

Tell us about it. We’ll size a first server and a first question worth answering.

Request a quote