# Acaysia

> Acaysia builds a control system that optimizes physical processes (yield, energy, throughput), proven first on chemical reactors. It runs on a typed compositional ontology of unit operations and uses MPPI (Model Predictive Path Integral) control, with PID fallback in under 100 milliseconds. Based in Oakland, California; founded 2025. Brand line: "Control the Impossible."

Acaysia runs on real continuous stirred-tank reactor (CSTR) hardware, not simulation. The safety architecture is ASIL-D inspired and SIL compatible and never interferes with Safety Instrumented Systems. Acaysia integrates over OPC UA and EtherNet/IP as a brownfield drop-in, compatible with Rockwell, Siemens, Beckhoff, Schneider, and ABB control systems. Acaysia is partnered with NVIDIA and is in a16z's accelerator.

## Core pages
- [Home](https://acaysia.com/index.md): Overview, FAQ, and an "at a glance" facts table.
- [Product](https://acaysia.com/product.md): The Acaysia MPPI control system.
- [About](https://acaysia.com/about.md): Company and technical approach.
- [Contact](https://acaysia.com/contact.md): Request a process control demo (founders@acaysia.com).

## Engines
- [AcaysiaRT](https://acaysia.com/engines/acaysia-rt.md): High-throughput GPU simulation runtime. One templated CUDA core serves 127 registered process models (plus 45 curated reaction-network processes), each bit-exact parity-tested against a PyTorch reference, with model-generic MPPI control. Measured ~6.1 billion reactor-steps/sec on one RTX 5080 (10,000 rollouts x H=100 in 0.164 ms), 182x the PyTorch reference. 92 of 127 models carry independent validation evidence at tiers T0-T2.
- [AcaysiaDRT](https://acaysia.com/engines/acaysia-drt.md): 3D spatial simulation engine extending AcaysiaRT with D3Q19 Lattice Boltzmann transport, for resolving what happens inside a unit rather than treating it as well-mixed.
- [AcaysiaGEM](https://acaysia.com/engines/acaysia-gem.md): Parametric geometry engine that authors the fluid domain AcaysiaDRT runs in - the negative space the fluid occupies, not the mechanical part. 18 parametric geometries across flow domains (6), inline pipe components (2) and full vessels (10: stirred tank, distillation column, bubble column, packed-bed reactor, fermenter, flash drum, decanter, cyclone, drum, horizontal separator). Boundary regions are tagged at build time, every domain is gated on watertightness and leak checks before export, and plants of many units can be placed, piped, and rendered photoreally or exported as 3D for the web.
- Rete (a layer of AcaysiaRT, documented at https://acaysia.com/engines/acaysia-rt.md): GPU-native recycling flowsheet engine. Wires registered AcaysiaRT models into a whole plant with recycle loops and marches the coupled plant batched over thousands of samples, with plant-wide MPPI on top. A 70% recycle raises overall conversion from 0.4623 to 0.5961 while per-pass conversion falls to 0.3068, cross-checked against DWSIM 9.0.5; whole-plant kernel fusion measures 22.07x over matched per-unit CUDA at 4 units (RTX 5080, N=1024, H=100).

## Platform
- [AcaysiaCORE](https://acaysia.com/core.md): Simulation-as-a-service - the unified analysis and exploration app for the engines above, in the browser. Status: in build, not yet generally available; early access by request at aias@acaysia.com. Explicitly NOT a control tool; the planned desktop OPC UA connector is read-only data acquisition for comparing plant behavior against simulation, with no write surface.

## Guides
- [What is advanced process control (APC)?](https://acaysia.com/resources/advanced-process-control.md): Definition, the PID -> MPC -> MPPI evolution, and where APC applies.
- [MPPI vs PID vs MPC](https://acaysia.com/resources/mppi-vs-pid-mpc.md): How path-integral control differs from PID and classical model predictive control.
- [Resources](https://acaysia.com/resources.md): Index of guides and documentation.

## Contact
- Email: founders@acaysia.com
- Location: Oakland, California, United States

## Machine access
- [llms-full.txt](https://acaysia.com/llms-full.txt): single-file Markdown corpus of every page.
- [skill.md](https://acaysia.com/skill.md): operating guidance for AI agents.
- MCP server at https://acaysia.com/mcp (Streamable HTTP, no auth; metadata at https://acaysia.com/.well-known/mcp.json).
- Every page listed above is served as Markdown via its .md URL variant or the Accept: text/markdown header.
- Additional technical documentation, case studies, and whitepapers under /resources/ are access-gated (HTTP 401 without an access code); request access at founders@acaysia.com.