Built by engineers who spent years on the plant floor watching avoidable failures happen.
In 2022, Thomas Osei was leading automation systems at a mid-sized stamping plant in Ohio. A compressor used in a critical press cooling loop failed on a Tuesday morning, shutting down three lines for eleven hours. The bearing had been showing textbook outer race wear patterns for weeks — measurable in the vibration data from the sensor that was already installed on the machine, just never analyzed.
The $340,000 in lost production was not a technology gap. The sensor was there. The data existed. What was missing was software that could run the analysis locally, continuously, and translate the result into a maintenance work order before the bearing seized.
EdgeRun AI was founded in Stamford in 2024 to build exactly that: an edge-native inference platform that connects to existing sensors, learns what normal looks like for each asset, and fires alerts into the maintenance system before the failure window closes.
Sending raw vibration data to the cloud is slow, expensive, and creates security exposure on operational networks. Every inference decision runs on-site, in milliseconds. Cloud connectivity is optional, not required.
A predictive maintenance alert that arrives as an email is one that gets missed. We push structured work orders into the CMMS the maintenance team already uses, during the shift where the information is actionable.
We publish our false positive and false negative rates per asset class and update them quarterly. A system that cries wolf gets switched off within a month. Our threshold-tuning process is designed to earn and keep operator trust.
Reduction in unplanned press stoppages across 38 monitored hydraulic press motors over a 6-month pilot. Maintenance cost per motor dropped from $4,200 to $2,400 annually.
Advance warning time before centrifugal pump bearing failures. Pilot covered 14 process pumps. Zero unplanned production stops during the 90-day evaluation period.
Estimated annual savings from two prevented motor failures at a water treatment station. Both events were detected 31 hours in advance, allowing scheduled repair during off-peak demand.
We run 90-day pilots with a fixed scope: one asset class, one gateway, and a clear ROI calculation at the end.