Acaysia: adaptive control for drifting industrial processes

Acaysia is building adaptive control software for physical processes, starting with chemical reactors. Its buyer is a plant operator that already has PLCs and advanced process control but still loses yield, energy or throughput as feedstock and conditions drift.

What it does

The current Acaysia site says the controller builds a process model from plant data and known physics, simulates thousands of trajectories with MPPI, optimizes yield, energy and throughput, and falls back to the existing PID in under 100 milliseconds when a fault occurs. The company says it runs on real reactor hardware, begins in shadow/advisory mode and retrains on-premises with versioned rollback.

Acaysia says its current proof comes from continuous stirred-tank reactors and publishes claims of more than 2% yield improvement, roughly 10% energy reduction and a failsafe response under 100ms. Those are company-reported operating results, not an independent plant-control benchmark. The site says the same ontology can extend to columns, dryers, crystallizers and evaporators, but the public proof is reactor-first.

Why I’d look closer

The advantage is brownfield compatibility. A plant does not need to replace its PLCs or safety instrumented systems to test a new optimization layer. Founder context fits the engineering problem: Achraf Zemzami worked on control systems and at Audible/AWS, Aias Tatsis is an electrical-engineering student focused on adaptive ML control, and Andy Xian studied statistics at Harvard.

The tradeoff is operational risk. “Failsafe” is a system claim that needs plant-specific validation, interlock review and a clear manual fallback. Process drift, sensor failure, model mismatch and abnormal feedstock can all make a controller optimize the wrong objective.

What I’d ask

Which sensors and PLC interfaces are required? Who approves the control envelope and model updates? How are shadow-mode predictions compared with the existing PID? Can operators see every proposed move, rollback a version and prove that the safety instrumented system remains independent? What evidence supports the reported yield and energy changes?

My editorial take

Shortlist Acaysia if a real process is leaving measurable yield or energy on the table and the plant can run a bounded shadow-mode pilot. Do not start with autonomous control. Prove model fidelity, fallback timing and operator trust on one unit operation before expanding the controller to adjacent processes.

Quick facts

Field	Sourced detail
Product	Physics-informed adaptive control for industrial processes
Buyer	Chemical, pharmaceutical and process-manufacturing plants
Current proof	Company says control runs on real CSTR hardware
Company claims	>2% yield improvement, ~10% energy reduction, <100ms PID fallback
Main question	Can the controller improve economics without weakening the plant’s safety envelope?

Sources checked

Source	Checked
Acaysia Speedrun profile	2026-09-19
Acaysia homepage	2026-09-19
Acaysia about page	2026-09-19