osrep.ai Request a site assessment

Know your machines are failing before they fail.

OSREP builds edge units that mount at the asset, read the sensors already wired there, and return one continuous machine state: nominal, degrading, or fault, with the input that drove the change. Inference runs on the unit. Raw data never leaves your site.

Request a site assessment How it works
The complete output vocabulary: nominal degrading fault If the state hasn't changed, there's nothing to do.

A pressure swing and a failing bearing can look the same.

Most monitoring treats normal as a fixed signature. It isn't. Normal moves with load, ambient temperature, and whatever the process is doing. A threshold set in January trips on a hot afternoon in August, the alarm says fault, and the machine is fine.

False alarms kill the tool

Operations gives a monitoring device a few chances. After the third alert that turns out to be a process change, every alert gets ignored, including the real one. A system that can't tell process from mechanics ends up unplugged.

Cloud models learn slowly, somewhere else

The usual fix is to ship waveforms offsite and let a model accumulate a season or more of history per asset. Your operating data lives behind a vendor's paywall while it learns, and stays there if you leave.

The data already exists. The review doesn't.

Historians hold years of vibration, acoustic, and process data that nobody has time to read at fleet scale. The failures are usually in there, visible in hindsight. Hindsight is the expensive kind.

Model normal as a function of conditions. What's left is mechanical.

An OSREP unit doesn't learn one baseline. It learns what the signal should look like given the operating context the other sensors report, then subtracts that prediction from what it observes. The residual is, by construction, the mechanical component. A process swing moves the prediction. A failing bearing moves the residual. The unit tells them apart because it never mixes them.

Step 1

Mount the unit

One unit at each monitoring point. It connects to whatever sensors are already wired there: vibration, acoustic, thermal, pressure, electrical. No rip and replace.

Step 2

Learn on the unit

A short commissioning window, on the hardware, at your site. The model starts from the physics of the equipment class instead of a blank slate, so this takes weeks, not a season of history. Nothing is transmitted while it learns.

Step 3

Emit a state

From then on the unit returns nominal, degrading, or fault, with confidence and time in state. When the state changes, it names the input that drove it.

The OSREP pipeline Five sensor inputs feed an edge unit. Inside the unit, the observed signal minus the expected signal for current conditions yields a residual, which is classified into a machine state. The example output shows a degrading state driven by acoustic input up 40 percent. sensor inputs (signal + operating context) vibration acoustic temperature pressure current OSREP edge unit observed signal minus expected signal for current conditions all inference on the unit residual output continuous machine state degrading driven by acoustic, up 40% temperature and vibration nominal
The pipeline. The unit predicts what the signal should look like under the conditions the sensors report, subtracts that from what it observes, and classifies the remainder. A process change moves the prediction. A mechanical fault moves the residual.

A state, not a stream.

Raw data stays on the unit. What travels upstream is already an answer.

Continuous machine state

Every unit emits a rolling state with a confidence value and time in state. It drops into your existing SCADA, historian, or operations platform over the protocols already on site: Modbus, HART, OPC-UA, MQTT. If nothing changed, nothing needs you.

Attribution with every change

A state change arrives with the input that drove it and by how much. Cause, not just symptom.

degrading: acoustic up 40%, temperature nominal, vibration nominal. That reads as a bearing starting to go.

Where this stands, plainly.

Early-stage hardware companies usually hide this section. We'd rather you read it here than discover it on the first call.

Hardware
Prototype built. On-unit inference is in active development on the bench in Baton Rouge. We are not shipping production units yet.
Certification
Not yet rated for Class I, Division 1 or 2 areas. That scopes current work to non-classified balance-of-plant and remote ancillary assets. Hazardous-area certification is on the roadmap, and we know it gates most classified facilities.
Early partners
Recruiting now, in two shapes. A data engagement that starts from a historical export out of your historian, so the model proves itself on your assets before any hardware ships. Or a small on-site deployment at non-classified monitoring points. Both start with the form below.
Company
OSREP LLC, founded 2025, Baton Rouge, Louisiana.

Who you'd be working with.

Garrett Wainright

Garrett Wainright

Chief executive, co-founder

Economics first, then a master's in computer science at LSU focused on running machine learning on constrained hardware, which is the problem OSREP exists to solve. Garrett reads every request that comes through this page.

The founding team includes nine industry and academic consultants and two technical co-founders, in cybersecurity and computer engineering. Garrett on LinkedIn.

Tell us what you're monitoring.

We'll reply with a short, specific assessment: whether OSREP fits your assets, what a deployment looks like at your site, and which of your existing sensors we'd work with.

What you'll get back. A reply from the founding team within two business days. No sequence, no list, no follow-up cadence.

Fields marked are required.