Imubit

Setpoints go straight into a live refinery control system on a cycle measured in minutes, an authenticated write with physical consequences and nothing advisory about it; the vendor positions explicitly against advisory only tools.

The counterparties are named by system class and protocol: "the on-premises closed loop optimization communicates with numerous on-prem sources, including the DCS, laboratory, and data historians, utilizing OPC/OPC-UA an industry standard protocol", and the application "can connect to either the DCS or to existing APC."

A fifth counterparty sits on the economics side, with prices pushed from the plant's LP planning system to the historian or straight to the DCS. That makes five distinct classes of industrial system.

OPC-UA is an open industrial standard, so the integration surface is defined by a published specification rather than a vendor maintained connector list. The cadence is published and configurable: "typically data is collected every 3 to 5 minutes. The range for all Imubit applications spans from a 1 minute read/write frequency to up to an hour." No control system vendors, such as Honeywell, Yokogawa, Emerson or Aspen, and no supported version matrix are published.

Sourceimubit.com/blog/answering-the-important-questions product and architecture FAQsread 2026-09-14

One controller does the work, though it runs a real multi step loop. Each cycle it reads live tags, checks them against a policy solved in advance, respects constraints, writes setpoints and watches the response, on a cycle of one minute to one hour. The behavior is conditional rather than linear: flexible MVs cover planned instrument outages, bad value protection comes "along with instructions of what to do in case of values in that range", and the model stops itself on an unplanned outage ("If it's unplanned, the model will turn itself off").

No second actor is named, no roles are distinguished and nothing coordinates a handoff. The vendor's more than 70 live DLPC applications, more than 20 on FCC units, are separate deployments on separate units, not collaborating actors in one workflow. There is no named runtime, workflow graph or authoring surface: a customer sets objectives, constraints and tuning parameters, not a sequence.

Sourceimubit.com/blog/answering-the-important-questions product architecture, application and continuous improvement FAQsread 2026-09-14

Each deployment is grounded in that plant's own data, but by training rather than retrieval. The foundation process model is a neural process simulator, "trained using a neural network rather than first principles", with reinforcement learning used "to train and pre-solve for all possible optimization solutions in an offline cloud environment (no online solver)". Plant knowledge is compiled into weights ahead of time, not retrieved at run time, and a trained controller is not an index, graph or embeddings layer.

What happens at run time is per cycle context assembly: the controller reads live tags from the DCS, laboratory and historians over OPC/OPC-UA, typically every 3 to 5 minutes and anywhere from one minute to an hour, and feeds current values to a pre-solved policy.

The grounding is real: the model is plant specific rather than generic, domain experts encode mass balance, gain tuning and the variable structure through AutoML, and economic context is fed live, with prices pushed from the plant's LP to the historian or DCS or entered by hand. The product has no retrieval architecture by design; what it keeps instead is a persistent causal model specific to each plant.

Sourceimubit.com/blog/answering-the-important-questions technology comparison and continuous improvement FAQsread 2026-09-14

Customers control the constraints on the controller, but no step has a person approve a setpoint before it commits. Domain experts "set the rules of the game", defining "which variables are independent / dependent, the economic objective, the time to steady-state, constraints", and keep live control of priorities, since "tunable parameters are introduced so that in real time, the user can update the relative importance of each objective."

Flexible MVs cover planned instrument outages, flexible CV constraints apply where a limit binds only part of the time, and bad value protection carries "instructions of what to do in case of values in that range". An automatic fail safe is documented: "If it's unplanned, the model will turn itself off." A controller that stops rather than acts on bad instrumentation is a real bound on autonomy.

The design point is the opposite of a gate: the vendor positions against advisory only tools and sells autonomous writing. Constraints bound what the agent may do but do not put a person between the decision and the action, and operator bypass after the fact is control, not approval before it.

Imubit articles on closed loop AI in manufacturing and on advanced process control appear to describe an advisory mode, where operators evaluate recommendations before closed loop is enabled, and a shadow mode that runs without automatic changes; whether either is an operating state a customer can select is not shown.

Sourceimubit.com/blog/answering-the-important-questions continuous improvement and people FAQsread 2026-09-14

A dated attestation is published, but no access controls a customer would manage. The data security policy, linked from every page footer beside a "Certified in" badge block, states: "Imubit has completed a SOC2 Type2 audit for the period of March 1st 2022 to February 28th 2023. The full report is available upon request and signing of an NDA." GDPR is covered by a readiness assessment.

The surrounding controls are specific: "Imubit encrypts all data both at rest and in motion", the cloud environments get continuous penetration tests, and a named information security management system sits under continuous monitoring. Employees and contractors take annual security awareness and data handling training, devices go through compliance processes, and backup and restore procedures are tested regularly.

Nothing names SSO, SAML, SCIM, OIDC or role based access control (RBAC), so a customer cannot learn who inside their own organization can do what, a gap that matters on a product that writes setpoints into a refinery's control system. The audit period ended February 2023 and no later period is claimed. The page gives a Houston address for Imubit, Inc. that differs from the site footer.

Sourceimubit.com/data-security-policy compliance and security sections, site footer Certified In blockread 2026-09-14

The customer's teams can track how the models perform, but no record of individual runs is documented. The cloud platform is "a collaborative environment for members of all teams to build, evaluate, and track performance of the closed loop models", which is observation of the agent's own work by the customer's teams.

Model interpretability is real but limited: a customer's account is that starting on a simple unit with "two MV handles which they were familiar with" let operators "see the results and understand them", because "building this initial trust is critical to moving on to more complex control paradigms where explainability gets trickier." Understanding why a model behaves as it does is not reconstructing a particular run.

There is no log of setpoint changes with the tag values and constraints behind each one, and no audit trail a compliance reader could walk back. A per write record likely exists in the customer's control historian, but that is the customer's system keeping its own record. "The Imubit team is continuously monitoring customer's DLPC applications and being proactive about suggesting retrains", which is the vendor watching its own fleet as service delivery rather than observability a buyer operates. The product page describes explainability dashboards that could carry a run level view.

Sourceimubit.com/blog/answering-the-important-questions product architecture, continuous improvement and people FAQsread 2026-09-14

No memory layer is documented. The controller improves by retraining, which bakes learning into model weights rather than retaining anything it reads back as context: the reinforcement learning controller is trained offline, and "the rate limiting step in retraining a DLPC model on a new operating regime is collecting enough data in that regime, typically a minimum of a couple of months", with retrains executed in two weeks or less and the vendor proposing them when operations shift. A model rebuilt this way replaces the old state rather than consulting it.

The site historian is the plant's own record and the application's data model, read for training rather than recalled as memory.

Inferential biasing, where "Imubit biases inferentials onsite using cumulative sum method or straight linear biasing method" to track recent lab results, is calibration of a signal, not context an agent reads to decide. No store, scope, lifetime, expiry or purge path is documented.

There is no conversational surface either, and what domain experts supply is configuration defined once at modeling time. None of this says anything about how sophisticated the controller is.

Sourceimubit.com/blog/answering-the-important-questions continuous improvement and product architecture FAQsread 2026-09-14

Two named on premises deployment options are published, with a choice between them: "There are two on-premises deployment options for DLPC in the DMZ and PCN." A buyer with a different requirement gets a different answer, the demilitarized zone or the process control network.

The component that acts is the one that runs on site: "Imubit's on-prem DLPC application can connect to either the DCS or to existing APC. The software sits on the PCN level in order to communicate with these systems." The closed loop controller writing setpoints is inside the plant network, not calling into it from a vendor cloud.

The data boundary is stated and unusually strong: "There are no two-way exchanges of data between the on-premises controller and the cloud database." For a refinery, a controller that does not phone home decides whether the product is deployable at all. The vendor's architecture answer splits the platform explicitly: a cloud based Industrial AI platform for building, evaluating and tracking models, and a separate on-premises closed loop application for execution, with training data extracted from the site historian, typically from the business network.

The cloud side runs on AWS and GCP "from multiple zones and regions", with no region a customer can select. No installation or architecture documentation describes what an on-premises deployment requires.

Sourceimubit.com/blog/answering-the-important-questions product and architecture FAQs, imubit.com/data-security-policy availability sectionread 2026-09-14

The named units are deployments, not assets a customer adopts. The vendor reports "more than 70 live DLPC applications running within our global customer base, more than 20 of these on FCC units", spanning preheat, fractionation, reactor/regen, C3/C4 splitter and gas plant units, with further experience on C2 and C3 splitters and steamcracker furnaces.

Each is an instance built for a particular unit at a particular plant and trained from that plant's own history ("the key is to ensure your training data set contains enough operating data for each of the different feed composition regimes"), so there is no starting artifact a customer adopts, only experience the vendor brings to building a new one.

A second named product line exists: the Imubit Industrial Solution Applications, AI tools for "operator training, real-time process monitoring, and planning tool augmentation", alongside the Optimizing Brain Solution. Neither is published as separately adoptable with its own entitlement or selection surface. The machinery behind them is real and named, from Optimizing Brain and Deep Learning Process Control to the Foundation Process Model.

Sourceimubit.com/blog/answering-the-important-questions application FAQs, imubit.com/product summaryread 2026-09-14

Work reaches Imubit's controller with nobody at a screen, on a published, customer configurable cycle: "This is a configurable parameter, but typically data is collected every 3 to 5 minutes. The range for all Imubit applications spans from a 1 minute read/write frequency to up to an hour." A process disturbance, a feed composition change, a price update or drift in unit conditions arrives as tag values and the controller responds; the vendor contrasts this with traditional optimization methods that "might update setpoints weekly or monthly, missing short-term opportunities".

External events are not all sensor data. Plants push prices "from their LP to their historian or straight to the DCS", and "Imubit can see in real time when there are changes to the model based on updates to the price deck", a cross system event trigger. Laboratory and analyzer results feed back to bias inferentials as they arrive.

Conditional suppression shows the trigger logic is real: "If it's unplanned, the model will turn itself off" on an instrument outage, with flexible MVs covering planned outages and bad value protection defining what to do in a bad range. No webhook, event subscription model or named scheduler interface is published.

Sourceimubit.com/blog/answering-the-important-questions product architecture and continuous improvement FAQsread 2026-09-14

Imubit runs its own proprietary architecture and the customer cannot choose models. "The dynamic process simulator is trained using a neural network rather than first principles", and reinforcement learning is used "to train and pre-solve for all possible optimization solutions in an offline cloud environment (no online solver)", which the vendor says "is unique to Imubit". No model selector, admin entitlement, per application model parameter, disclosed routing, or bring your own model or key path appears anywhere.

What the customer controls is the objective, not the model: domain experts define independent and dependent variables, the economic objective, time to steady state and constraints, AutoML translates that into model parameters, and the relative importance of each objective is tunable in real time. Steering what a model optimizes for is not choosing which model runs.

The CEO's launch quotation that "ChatGPT and other foundation models are reshaping industries with their broad applicability" frames the term foundation process model and does not name an engine. No third party model provider is named; AWS and GCP are compute, not models.

Sourceimubit.com/blog/answering-the-important-questions technology comparison FAQs, imubit.com/data-security-policy infrastructure sectionread 2026-09-14

No interface for calling the platform is published: no REST API, SDK or MCP server appears anywhere in the site navigation or footer. There is no developers entry and no documentation site, and insight.imubit.com hosts a value assessment tool used for lead capture rather than a developer surface.

The integration story is strong but points entirely outward. OPC/OPC-UA connections to the DCS, laboratory systems and data historians, and setpoint writes into the customer's control infrastructure, are the platform reaching into other systems; OPC-UA is how the vendor drives the plant, not how an outside caller drives the vendor. The cloud Industrial AI platform lets a customer's teams build, evaluate and track closed loop models, but an application a person logs into is not an extension surface.

Imubit makes no API claim at all, which fits what it sells: a delivered outcome on a production license, implemented and maintained by its own engineering team, with a customer describing that "a lot of the supporting work to implement and maintain would be done by the Imubit team." An enterprise deployment may expose interfaces under contract that nothing published describes.

Sourceimubit.com navigation and footer read across three pages, imubit.com/blog/answering-the-important-questions product architecture FAQsread 2026-09-14

Customers get an evaluation surface of their own and a retrain loop they review. The cloud Industrial AI platform provides "a collaborative environment for members of all teams to build, evaluate, and track performance of the closed loop models": evaluation is a named function of a platform the customer's teams use, and the object is the agent's own work, this customer's controller on this customer's unit, not a benchmark or the plant's health.

A post deployment loop runs alongside it: "the Imubit team is continuously monitoring customer's DLPC applications, and being proactive about suggesting retrains when they see shifts in operation", with customers brought in "to review updates to the model following a retrain." Detecting drift, proposing a change and reviewing the changed model with the customer before it runs is a gate in a release path.

Retraining is time boxed: a new operating regime needs a couple of months of data and two weeks or less to execute. The offline reinforcement learning step is how the model is built, not how it is evaluated. No scored test cases, golden set, accuracy figure or published method for what evaluate means inside the platform is documented, and the review step is run by the vendor's team rather than invoked by the customer.

Sourceimubit.com/blog/answering-the-important-questions product architecture, continuous improvement and technology comparison FAQsread 2026-09-14

Plant control systems are reached through a published industrial protocol, not through a browser or computer interface. The on-premises controller "communicates with numerous on-prem sources, including the DCS, laboratory, and data historians, utilizing OPC/OPC-UA an industry standard protocol", and writes setpoints through that channel on a one minute to one hour cycle. OPC-UA is a machine to machine specification: there is no screen in the loop, no cursor and no rendered control to target, and no hosted or local browser, desktop session or remote computer control is involved.

How often it breaks when UI elements change does not arise, because OPC-UA tags are a stable contract and the stated failure mode is instrument outage rather than interface drift, with the model turning itself off. The cloud Industrial AI platform is an application the customer's teams use, so a person operates that screen. The architecture is described precisely enough that a screen driving path is implausible: a system that drove a DCS operator console instead of writing tags would be a different product and a different safety case.

Sourceimubit.com/blog/answering-the-important-questions product and architecture FAQsread 2026-09-14