AI that runs the equipment, not just the dashboard.
Models deployed at the edge, on premises and air‑gapped if you need it. From supervisory setpoints down to driving the equipment directly, with deep reinforcement learning where it counts.
Driving the valve, not just the setpoint.
Since 2020 our agents have driven the final control element itself, not a setpoint for the PID to chase, on the plant's own controller with the PID tracking underneath.
- 2019
Founded by controls engineers
Engineers, data scientists and industry veterans who saw a gap between industrial control systems and modern AI.
- 2020
Deep RL on the valve
A deep Q-network took a refinery separator's outlet valve from its PID, on the plant's own Allen-Bradley PLC. To our knowledge, no one had given deep RL that seat on a live plant before. It still runs that valve today.
- Now
What we've been up to since
SAC and PPO in running plants. Agents trained with soft actor-critic and proximal policy optimization run loops on the PLCs and DCSs those plants already had.
Across industries. Refinery separators and blend headers, chemical reactors, powder fill lines and commercial refrigeration, just a few of the industries we've been optimizing.
Your own Python on live plant data. Koios components let your data scientists run their own Python beside our models, on the same live plant data.
We've been busy proving deep RL has a home in any part of controls, from a single valve to an entire process.
From a forecast on a screen to a hand on the valve.
Most AI in a plant stops at a dashboard. We work all the way down the stack, and the closer it gets to the equipment, the more our experience matters.
- Adviseoutput: alerts and trends
Forecasting and anomaly detection
What's coming, and what looks wrong, before an operator would see it.
- Inferoutput: a live estimate
Soft sensors
The values you can't measure, inferred live from the ones you can.
- Superviseoutput: setpoints
Supervisory control
The model writes setpoints to the loops you already run.
- Actoutput: the equipment
Base-level control
The model drives the end equipment directly, not just a setpoint. This is where our deep reinforcement learning work lives.
Deep reinforcement learning, in control of real equipment.
We train a controller, not just a predictor. It learns from your own process history, proves itself in shadow, then takes over inside limits you set.
- 01
Learn from your history
Trained against the data your historian already holds.
- 02
Run in shadow
It watches the live loop and proposes moves no one acts on yet.
- 03
Take over inside limits
Once it controls the loop better than what you run today, it takes over, inside limits you set.
Same Koios, whichever box you give it.
Models, Python and control logic move between hosts unchanged. Pick the one that sits closest to the loop it serves.
Ubuntu Server and Docker on every one. The installer carries both, so no internet is needed.
Train it, run it, ask about it.
Koios
Runs your models, your Python and your control logic next to the equipment, on edge hardware, a server or a VM.
- One container, on bare metal or a VM
- Talks OPC UA, Modbus TCP and EtherNet/IP, SLC 500 and MicroLogix included
- Runs your own Python: MPC, optimization, anything
Ronin
Training made simple for classification, soft sensors, forecasting and anomaly detection.
- Clean, trend and train in one place
- Graded on history it never saw
- Export to ONNX and run it in Koios
Curtis
Plant insight and a partner to think out loud with. An assistant that knows your Ronin projects and your account in the portal.
- Answers from your own projects
- Brainstorms which tags to try next
- Plain answers, with sources
Built for processes that can't stop.
The dryer is tuned to the worst case.
Moisture is checked at the end of the line, so every correction is for product that has already been made. The setpoint runs hot, and the energy bill pays for the safety margin.
Infers outlet moisture continuously from inlet conditions and line speed, then trims the setpoint before the deviation reaches the product.
Bring us your most difficult process.
The nonlinear one, the one that keeps drifting, the one PID never quite holds. Reach out with any questions, or ask about a 30-day trial.