Industrial
Manufacturing, logistics, ports and infrastructure: environments where the data is partial, the systems are old, and the tool has to work the first time.
Industrial operations are where AI meets its least forgiving conditions. Readings arrive incomplete, the system of record was chosen long before anyone planned for AI, and the person who has to trust the output is a dispatcher, a planner or a technician standing in front of a machine. We build for those conditions rather than around them, starting from the data that actually exists and the systems already running the operation.
Partial data is the starting condition
Industrial estates produce readings, scans, tickets, forms and free text, and almost none of it was collected with a model in mind. Sensors drop out. Fields are left blank because the shift was busy. The same asset is named differently in every system that touches it. We treat that as the design input rather than a cleanup project that has to finish first, building pipelines that reconcile what exists and systems that say plainly when the evidence is too thin to answer.
The estate already running the operation sets the terms
Nothing in an industrial operation gets replaced because it would suit an AI project. The scheduling system, the maintenance record, the warehouse and transport platforms and the process historian all stay, and anything we build has to work through them. So we design the integration layer first and the intelligence second, taking read access before write access and putting an audit trail behind every action a system is allowed to take. The founders run a live network and security engineering firm across multi-vendor estates, and separately built an attack-surface management platform, so integration and exposure get designed as one question rather than two.
The back office runs on documents
Much of an industrial operation still runs on paper and its digital descendants: bills of lading, certificates, inspection reports, work orders, warranty and supplier claims, each one filed somewhere a person has to go and find. This is the least glamorous AI work and often the most valuable, because those documents are already the record the business argues from. We build extraction and retrieval over them with the provenance intact, so an answer can be traced back to the page it came from and defended when someone disputes it.
Operators decide whether the system survives
A dispatcher or a maintenance technician will abandon a tool the first time it is confidently wrong, and it never gets a second hearing. That is why evaluation is written before the build starts, why every answer carries its source, and why the system is designed to hand the decision back to a person when it is out of its depth. Training and runbooks for the people on the floor are deliverables in the engagement, not a slide at the end of it.
What we build in these environments
- Data pipelines that get sensor, telemetry and work-order data out of the systems holding it and into a form both a model and a planner can use.
- Condition monitoring and maintenance decision support that shows the readings behind every flag it raises.
- Planning and dispatch support for logistics and field operations, built against the constraints the operation actually runs under.
- Extraction and retrieval across shipping, inspection, warranty and compliance documents, with every answer traceable to its source page.
- Evaluation suites, runbooks and operator training, so the people on the ground can judge the system and keep it honest.
What this is not
This is not control engineering. We do not program PLCs, commission line equipment or touch safety instrumented systems, and we do not sell a manufacturing platform of our own. Caltor is also not a staffed industrial practice. It is a small firm whose founders have built and operated systems under these conditions, and who take on industrial work where that engineering and security discipline transfers. Where it does not, we say so in the first conversation.
Talk to a partner
Tell us which part of the operation is still running on guesswork, and what the data behind it actually looks like.
