My path into AI did not come from a CS PhD or a hyperscaler internship. It came from oil and gas: refineries, LNG terminals, and the kind of operations where systems must work, data is sensitive, and a single hour of downtime can cost millions.
Reliability engineering teaches a discipline most of the AI world has not had to develop yet. You learn to respect ugly interfaces, partial information, aging assets, bad labels, hidden tribal knowledge, and the operational cost of being wrong. The job is not to sound smart in a meeting. The job is to keep the system online.
The arc
I started close to the equipment: rotating machinery, functional safety, inspections, maintenance strategy, and the CMMS workflows that turn intent into work. Over time, the pattern became hard to ignore. The best operators were drowning in useful data, but the tools around them were still built for record keeping, not judgment.
That is the thread that led to Livit AI. Heavy industry does not need vague AI transformation. It needs systems that can live inside real constraints: private data, on-prem environments, high consequence decisions, and people who already know where the process is brittle.
The thesis
The next useful wave of industrial AI will not be won by generic models alone. It will come from pairing domain context with infrastructure that can actually be deployed where the work happens. That means local inference, retrieval that respects messy source systems, workflows that earn trust, and hardware decisions that make the economics possible.
I am especially interested in the overlap between reliability engineering and LLM operations. Both fields care about failure modes, drift, observability, maintainability, and what happens after the demo is over. The industrial world has a lot to teach AI builders about systems that have to keep working.
The personal
Outside Livit AI, I tend to stay near hardware and edge cases. The robotics tinkering is partly a fabrication project and partly a reminder that software always ends up negotiating with geometry, sensors, power, and the physical world. The local LLM project is a public place to work through tooling, model deployment, and the practical limits of running capable systems on machines people can actually own.
I like talking to industrial operators about real AI deployments, hardware people working on PCIe and GPU systems, and builders making on-prem AI tooling sharper. Those are usually the conversations worth having.