Expertise

Applied artificial intelligence and systems integration

Artificial intelligence that stays in production every day, on the processes that actually weigh. And systems that finally talk to each other, without throwing away what works.

These two live on one page because inside a company they are the same thing. An AI assistant that cannot see the real data is a demo. A model that reads documents but cannot attach them to the management system has moved the work, not removed it. Useful artificial intelligence is almost always an integration problem in disguise: the value is not in the model, it is in the pipe that carries the data in and the results out.

What I put into production

• Documents that read and file themselves. Capture at the point where the document is born, classification and extraction with a model, attachment to the management system data, indexed filing. With a human review queue for the cases where the model is not confident, because that is what keeps the automation standing over time.
• Assistants that answer on company data. Questions in plain language, answers built by running queries against the real records, with the permissions of whoever is asking and read-only access.
• Agents that run processes. Not only answering: doing. A repetitive step spanning three systems becomes a procedure that runs on its own and asks for help only when it needs to.
• Integration between systems. The legacy management system, the cloud, external services, the spreadsheets nobody dares touch. Making what you have talk almost always costs less than replacing it.

The principle: the emergency exit

An automation with no plan for the human exception gets abandoned at the first strange case, and rightly so. That is why every system I build has a point where, when the model is not confident, it does not guess: it queues the case and asks. A person looks, confirms or corrects, and the system learns. This is not a patch: it is the reason these things have been in production for years rather than for a quarter.

The second principle concerns the data: it stays where it is, with the permissions it has. AI adds an interface on top of the same data, for the same people, under contracts and configurations that exclude the use of your data for training models.

Related reading: A hundred and ten documents a day, untouched · Asking your management system questions in plain language

How it starts

From a single process, the one that hurts most, chosen together by looking at where hours and errors pile up. The first weeks have to return something usable already: if an AI project needs six months before showing a result, the shape of the project is wrong, not the technology.