Archimetis vs Imubit (2026)

Plant data comes in by system class, in Archimetis's words "Connect DCS/historian data, integrity systems, incident logs, and process safety studies into a single model of your plant." A DCS is the plant control system and a historian is the time series system of record, both below the application layer. Document classes are named too (Bowtie, LOPA and PHA studies, manuals, historical reports and incident notes). The worked examples reference live data by instrument tag (EI-401, AI-101, TIC-7501.SP, K7502A), which points to a live historian connection.

Archimetis describes the transport only by what it does not need. "Archimetis does not require software agents, inbound firewall rules, VPN tunnels, or changes to your OT network. Data is shared through customer-approved, outbound-only, cloud-native mechanisms." Customers learn what will not be asked of their network but not how the connection is made. No protocol is named, whether OPC-UA, MQTT, a historian API, a file drop or a message bus, and there is no named product or connector catalog.

Archimetis only reads. It describes no writes into a DCS, a historian or a maintenance system, and the product is advisory by design.

Sourcearchimetis.ai/refining Integrate your plant and Enterprise-Grade Security card 03, /chemicalsread 2026-09-13

Setpoints go straight into a live refinery control system on a cycle measured in minutes. Each one is an authenticated write with physical consequences, and Imubit positions itself against tools that only advise.

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". The application "can connect to either the DCS or to existing APC." On the economics side, prices are 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 connections follow a published specification instead of a connector list Imubit maintains. The cycle is configurable. Data is typically collected every 3 to 5 minutes, and the read and write frequency across Imubit applications ranges from 1 minute to up to an hour. Imubit names no control system vendors, such as Honeywell, Yokogawa, Emerson or Aspen, and publishes no supported version matrix.

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

A five step sequence runs unattended, but it is Archimetis's fixed method and customers cannot compose their own.

The method reads the plant's existing Bowtie and LOPA documentation to find the highest consequence scenarios, checks each against live and historical operating data, and evaluates whether safety barriers are stressed during upsets, "using control performance, not alarms or trips."

It then turns multisignal trends into an operating picture and surfaces the most likely risk signature and the next best checks. In the worked example, Archimetis searches documentation, extracts ranked scenarios, selects parameters, plots a month of tag data and interprets the margin. It then "formulates and ranks hypotheses, validates them against process data."

Archimetis owns every step. It names no workflow builder, step editor, exposed branching or conditional logic, or orchestration component, and customers cannot author, save, version or reuse a flow. The nearest customer control is "Engineers define the decision space. They set constraints, validate assumptions," which sets parameters for an analysis Archimetis designed.

Sourcearchimetis.ai/refining five-step How-to module and worked HDS example, /chemicalsread 2026-09-13

A single controller does the work in a loop. Each cycle, it reads live tags, checks them against a policy solved in advance, respects constraints, writes setpoints and watches the response. The cycle runs from one minute to one hour. The loop changes with conditions. Flexible MVs cover planned instrument outages, and bad value protection comes "along with instructions of what to do in case of values in that range". On an unplanned outage the model stops itself ("If it's unplanned, the model will turn itself off").

Imubit names no second actor, no distinct roles and no handoff. Its more than 70 live DLPC applications, more than 20 of them on FCC units, are separate deployments on separate units, each working on its own. Imubit names no runtime, workflow graph or authoring tool. 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

Detection runs overnight, unattended. "Unusual reactor conditions are detected over nightshift. Archimetis identifies possible causes and begins to triage and prioritize the actions required for operations. Engineers are in control, even before their morning meeting." The work runs with nobody logged in, and the results are waiting when engineers arrive.

Four kinds of trigger are named. The schedule is the shift, with "better decisions, from the first minute of every shift" and start of shift as a named state. Data conditions fire "when the data looks alarming," with timestamps such as "At 17:45, the WABT was detected to exceed its allowed operating window. This continued into nightshift." Forecast thresholds fire on a projection instead of a reading, such as "Fouling rate forecast: E4405 will exceed limit in 14 days," with the normalized fouling rate accelerating toward a 9.0 bar limit. Calendar state fires "when turnarounds loom."

Findings carry clock times in the product UI (7:01 AM, 9:05 AM), with severity High and ownership routed to Panel, Field Operations, Outside Operator or Maintenance. They arrive as detections that have already fired, not a dashboard a person has to open.

Archimetis also says "Rather than issuing alerts, Archimetis frames the next best investigation." That describes the form of the output, a framed investigation instead of a bare alarm, within the same five step sequence that begins with continuous checking against live data.

Archimetis names no trigger configuration surface, webhook, external event subscription or schedule editor. Archimetis sets the triggers, not the customer.

Sourcearchimetis.ai/refining hero scenarios and How-to section, /chemicalsread 2026-09-13

The controller runs on a cycle the customer configures, with nobody at a screen. Data is typically collected every 3 to 5 minutes, and the read and write frequency across Imubit applications ranges from 1 minute 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. Imubit contrasts this with traditional optimization methods that "might update setpoints weekly or monthly, missing short-term opportunities".

Plants push prices "from their LP to their historian or straight to the DCS", and Imubit sees in real time when updates to the price deck change the model. Laboratory and analyzer results feed back to bias inferentials as they arrive. On an unplanned instrument outage the model turns itself off. Flexible MVs cover planned outages, and bad value protection defines what to do in a bad range. Imubit names no webhook, event subscription model or scheduler interface.

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

Reasoning is grounded in a maintained model of the customer's plant. Archimetis connects "DCS/historian data, integrity systems, incident logs, and process safety studies into a single model of your plant." The model is built once and reasoned over repeatedly, under a step called "Integrate your plant."

Archimetis "ingests plant documentation, manuals, historical reports, and incident notes," and names the formal document classes as Bowtie, LOPA and PHA studies, which make up the plant's approved risk framework.

Each conclusion links back to its sources. Archimetis says "Reasoning is fully traceable. Every conclusion can be traced back to live data and plant documentation, making recommendations easy to inspect, challenge, and refine." The reasoning runs "using the plant's approved risk framework, not a generic AI model."

Retrieval is a discrete step. In the refining worked example the agent says "I'll search our documentation for formal risk assessments, such as Bowtie or LOPA studies, that specifically cover the HDS unit," followed by a progress marker reading "Searching for HDS risk assessment documents." The answers cite named equipment tags and documented consequence bands from those studies, so findings come back attributed to their sources.

Archimetis does not disclose its retrieval technology, such as an index or embeddings.

Sourcearchimetis.ai/refining and /chemicals How we turn your data and expertise into decisions, Steer decisions with human judgmentread 2026-09-13

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

At run time, each cycle the controller reads live tags from the DCS, laboratory and historians over OPC/OPC-UA and feeds current values to the pre-solved policy. That happens typically every 3 to 5 minutes and anywhere from one minute to an hour. The model is specific to each plant, not generic.

Domain experts encode mass balance, gain tuning and the variable structure through AutoML. Economic context is fed live, with prices pushed from the plant's LP to the historian or DCS or entered by hand. What the product keeps is a persistent causal model for each plant, with no retrieval architecture.

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

In Archimetis's words, the product "turns ingested plant documentation and history into a living system that learns from every shift, decision, and outcome to preserve and scale institutional expertise." The ingested documentation is the corpus the system reads. The learning is a claim that the model improves over time, with no named store, scope, read or write surface, or lifetime. Preserving institutional expertise means knowledge that used to live in a retiring engineer's head now lives in the corpus.

Archimetis names no memory scope, lifetime or purge path. It describes no cross session recall, conversation history, saved context or preference store, and no case of the agent referring back to a prior run. The worked examples begin cold each time and set the scope again by asking the engineer "Which path seems best to you."

The plant model itself persists as an application data model built from customer documents.

Sourcearchimetis.ai/refining and /chemicals Learn from experienceread 2026-09-13

The controller improves by retraining, which builds learning into the model's weights. It keeps nothing to read back as context. The reinforcement learning controller is trained offline. Collecting enough data in a new operating regime is the slowest step in retraining a DLPC model and typically takes a minimum of a couple of months. Retrains are executed in two weeks or less, and Imubit proposes them when operations shift. Each retrained model replaces the old state.

The site historian is the plant's own record and the application's data model, and it is read for training. Imubit "biases inferentials onsite using cumulative sum method or straight linear biasing method" to track recent lab results, which calibrates a signal. Imubit names no memory store, scope, lifetime, expiry or purge path. There is no conversational surface, and what domain experts supply is configuration defined once at modeling time.

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

Engineers steer the analysis before the system reasons. Archimetis frames this as steering decisions with human judgment. "Engineers define the decision space. They set constraints, validate assumptions, and inject real operating context into the analysis." A person bounds the space, approves the assumptions and supplies context.

The output goes to review first. Archimetis calls its approach "AI + human expertise (not AI instead of humans)," and says it "turns data into action, showing the reasoning behind every recommendation so engineers stay in control." The system produces "ready-to-act recommendations," which a person then acts on. Nothing reaches the plant without a person.

The agent asks for a choice at each branch, as the worked examples show. It asks "Which path seems best to you," offering a plant wide search, a focused unit investigation or a specific hazard analysis. It then asks "Which of these risks would you like to investigate further," and, after ranking three hypotheses, "Which of these paths would you like to explore first." The fifth method step says the same thing, that "Rather than issuing alerts, Archimetis frames the next best investigation... so engineers stay in control."

There is no approval queue, escalation path, risk tiering per action or sign off record, because the system takes no autonomous action and never touches the controls.

Sourcearchimetis.ai/refining Steer decisions with human judgment, worked HDS example, How-to step 05, /chemicalsread 2026-09-13

Customers set the constraints on the controller, and 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". They 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, and flexible CV constraints apply where a limit binds only part of the time. Bad value protection carries instructions for what to do with values in a bad range. An automatic fail safe stops the controller on an unplanned outage ("If it's unplanned, the model will turn itself off").

Imubit sells autonomous writing and positions itself against tools that only advise. Constraints bound what the controller may do, but no person stands between the decision and the action. Operators can bypass the controller after the fact.

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

Customers sign in through "SSO/OIDC identity integration," and Archimetis adds "least-privilege service accounts" and "no standing credentials." It promises "Zero-trust access that satisfies enterprise IT."

Archimetis names no certificate, auditor or report, whether SOC 2, ISO 27001 or a penetration test summary, and has no trust center. It says only that it meets "strict enterprise compliance expectations," without naming a standard or anyone who verified it. Single sign on through OIDC is the one control customers use, and Archimetis names no role based access, SCIM or permission model over the agent.

Each customer runs in an isolated GCP environment, with TLS 1.3 in transit and KMS managed encryption at rest. Archimetis needs "no software agents, no inbound firewall rules, no VPN tunnels, no changes to your OT network," and data moves through "customer-approved, outbound-only, cloud-native mechanisms." These are properties of how the service is built.

Sourcearchimetis.ai/refining and /chemicals Enterprise-Grade Security Built Inread 2026-09-13

Imubit has completed a SOC 2 Type 2 audit, and the full report is available on request after signing an NDA. GDPR is covered by a readiness assessment. Imubit encrypts all data both at rest and in motion, its 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.

Imubit names no SSO, SAML, SCIM, OIDC or role based access control (RBAC) for the people in a customer's organization.

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

Audit logs cover every action, in Archimetis's words "complete audit logs for every action." Its actions are agent actions, so that log records what the agent did.

A user can also follow a run step by step in the product, as the worked example shows. The agent states its plan ("I will plot the following key parameters for the reactor over the past month," then five named tags) and emits progress markers ("Searching for HDS risk assessment documents," "Assessing reactor temperature performance against safety limits").

It renders the chart it reasoned over and reports what it found, including the negative result, "While we haven't had an actual runaway event, operating this close to the safety limit reduces our ability to respond." A user can see which documents were searched, which tags were pulled, over what window, and how the conclusion followed.

Archimetis describes no run history view, activity log a customer opens, trace export or admin console. It states no retention period and does not say who can read the audit logs.

Sourcearchimetis.ai/refining Enterprise-Grade Security card 02 and worked HDS example, /chemicalsread 2026-09-13

The customer's teams track how the models perform in the cloud platform, "a collaborative environment for members of all teams to build, evaluate, and track performance of the closed loop models". One customer describes starting on a simple unit with "two MV handles which they were familiar with", which let operators "see the results and understand them". In that customer's words, "building this initial trust is critical to moving on to more complex control paradigms where explainability gets trickier."

Imubit names no log of setpoint changes with the tag values and constraints behind each one, and no audit trail a compliance reader could walk back. The Imubit team continuously monitors customers' DLPC applications and suggests retrains as part of its service. Imubit also offers explainability dashboards.

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

Each customer runs in a fully isolated GCP environment with VPC peered transfers, so operational data is never shared or mixed. Archimetis names no region and makes no residency commitment. There is no self hosted option, and customers have no published way to choose where their environment runs.

GCP is the cloud Archimetis itself runs on. VPC peered transfers are a private network path between clouds, which changes how traffic reaches a tenant but not where the tenant is. Archimetis installs nothing in the customer's plant, with "No software agents, no inbound firewall rules, no VPN tunnels, no changes to your OT network," so there is no customer hosted option either. TLS 1.3 and KMS managed encryption at rest protect the data in transit and at rest.

Sourcearchimetis.ai/refining Enterprise-Grade Security cards 01 and 03, /chemicalsread 2026-09-13

DLPC has two on premises deployment options, the demilitarized zone (DMZ) or the process control network (PCN), and the customer chooses between them. The part that acts runs on site. The on premises DLPC application connects to either the DCS or existing APC and sits at PCN level to communicate with those systems. The closed loop controller that writes setpoints is inside the plant network and does not call in from a vendor cloud.

Data does not flow both ways between the on premises controller and the cloud database. The platform has two parts. A cloud based Industrial AI platform is for building, evaluating and tracking models, and a separate on premises closed loop application handles execution. Training data is 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. Imubit publishes no installation or architecture guide for an on premises deployment.

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

One system covers many process types. Archimetis "adapts to a wide range of processes," with one line each for Bulk Chemicals, Ammonia, Petrochemicals, Specialty Chemicals, Polymers and Fine & Specialty on what it reasons across. Those lines name reforming, heat recovery, compression and synthesis. They also name distillation columns, crackers and reactors, polymerization reactors and downstream processing, and batch and continuous campaigns. Customers do not select among the six.

Ammonia, for example, is a paragraph saying the same reasoning engine understands ammonia synthesis, not a unit a customer adopts.

Archimetis has no page or URL per item, no agent names, and no selection or enablement surface, marketplace, library, template gallery, install action or entitlement language. The industry pages (/refining, /chemicals, /industrials) market one product, Archimetis, which reads, flags and recommends. The four benefit cards on the refining page (Smarter Process Safety, Continuous Cost Optimization, Reliability Informed by Insight, Economic Performance at Full Potential) are outcomes the one system delivers.

Sourcearchimetis.ai/chemicals Archimetis adapts to a wide range of processes, /refining Four ways we improve how your plant runsread 2026-09-13

More than 70 live DLPC applications run across Imubit's global customer base, more than 20 of them on FCC units. They span preheat, fractionation, reactor/regen, C3/C4 splitter and gas plant units, with further experience on C2 and C3 splitters and steamcracker furnaces.

Each is built for a particular unit at a particular plant and trained from that plant's own history. In Imubit's words, "the key is to ensure your training data set contains enough operating data for each of the different feed composition regimes". Customers do not adopt a ready made starting point, and Imubit's team brings its experience to building each new one.

A second product line, the Imubit Industrial Solution Applications, offers AI tools for "operator training, real-time process monitoring, and planning tool augmentation", alongside the Optimizing Brain Solution. Imubit names no separate entitlement for either one. The named machinery behind them runs 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

The model is not the customer's to choose. Archimetis names no model selector, admin entitlement over models, choice per request or per agent, routing policy, or bring your own model or key path, and it sells no tiers.

No provider or model family is named. Archimetis says only "advanced AI," "AI that understands your plant" and "an AI partner." The one line on the subject is about grounding, "using the plant's approved risk framework, not a generic AI model," which ties the reasoning to customer documentation and does not let a customer pick the engine.

Sourcearchimetis.ai/refining, /chemicals, homepageread 2026-09-13

Imubit runs its own proprietary architecture. "The dynamic process simulator is trained using a neural network rather than first principles", and reinforcement learning is used to pre-solve for all possible optimization solutions in an offline cloud environment, with no online solver, an approach Imubit says "is unique to Imubit". It names no model selector, admin entitlement, per application model parameter, routing, or bring your own model or key path.

The customer controls the objective, not the model. Domain experts define independent and dependent variables, the economic objective, time to steady state and constraints, and AutoML translates that into model parameters. The relative importance of each objective is tunable in real time. 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 names no engine. No third party model provider is named, and AWS and GCP supply compute.

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

No developer surface is published. There is no docs or developer site, developer portal, API or GraphQL reference, SDK, CLI, MCP server, A2A card, webhook publisher, marketplace or app listing, and no partner or integrations program page. Demos are requested at /request-demo.

Customers cannot extend the product either. Engineers "define the decision space. They set constraints, validate assumptions, and inject real operating context," which sets parameters for an analysis Archimetis designed. Customers cannot add a connector, author an agent, define a trigger, or call the platform from outside.

Data flows in from the customer's side through "outbound-only, cloud-native mechanisms," and recommendations go back to people. No external system can command Archimetis, query its state or build on it.

Sourcearchimetis.ai homepage and site-wide footer, /refining, /chemicalsread 2026-09-13

There is no published REST API, SDK or MCP server for calling the platform, no developers entry and no documentation site. The insight.imubit.com site hosts a value assessment tool used for lead capture.

OPC/OPC-UA connections to the DCS, laboratory systems and data historians, and setpoint writes into the customer's control infrastructure, reach out from the platform into other systems. OPC-UA is how Imubit drives the plant, not how an outside caller drives Imubit. The cloud Industrial AI platform lets a customer's teams build, evaluate and track closed loop models, and people log in to use it.

Imubit sells a delivered outcome on a production license, implemented and maintained by its own engineering team. One customer says "a lot of the supporting work to implement and maintain would be done by the Imubit team", and Imubit makes no API claim at all.

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

No way to test a change to the agent is described. Archimetis names no harness before deployment, scored test cases, regression suite, golden set, A/B or shadow comparison of agent configurations, judge or rubric, or accuracy, precision or hallucination metric, and it publishes no quality figure.

The product itself continuously evaluates a refinery. It assesses whether safety barriers are being stressed, validates hypotheses against process data, ranks causes by likelihood, and forecasts catalyst life. "Continuous cost optimization" and "the platform continuously analyzes energy use and process trade-offs" describe an optimization loop over the customer's plant, not over the agent.

Every suggestion "can be challenged and improved," which is human review of each output. The system "learns from every shift, decision, and outcome," with no mechanism named. When the agent "validates them against process data," it is checking its own hypothesis about the plant during a run. The demo includes a "custom ROI analysis for your operations," which people produce during presales.

Sourcearchimetis.ai/refining Benefits and How-to sections, /chemicalsread 2026-09-13

The customer's teams evaluate their own closed loop models in the cloud Industrial AI platform, "a collaborative environment for members of all teams to build, evaluate, and track performance of the closed loop models". What they evaluate is their own controller on their own unit.

After deployment, "the Imubit team is continuously monitoring customer's DLPC applications, and being proactive about suggesting retrains when they see shifts in operation". Customers are brought in "to review updates to the model following a retrain", so a changed model is reviewed with the customer before it runs.

A new operating regime needs a couple of months of data, and a retrain takes two weeks or less. The offline reinforcement learning step is how the model is built. Imubit names no scored test cases, golden set or accuracy figure and does not say what evaluate means inside the platform. Imubit's team runs the review step, and the customer does not start it.

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

Every connection is a data path into a system of record, not a screen. "Data is shared through customer-approved, outbound-only, cloud-native mechanisms," and the connection is to "DCS/historian data, integrity systems, incident logs, and process safety studies." Archimetis describes no browser session the agent drives, no desktop control and no remote or local computer control.

A DCS vendor could redesign its operator console entirely and Archimetis would be unaffected, because it never sees the console.

Archimetis also needs no "software agents, inbound firewall rules, VPN tunnels, or changes to your OT network," so nothing of its own runs on the customer's endpoints.

The product's own chat and dashboard is Archimetis's client, not an interface it operates for someone. The product replaces a person reading screens ("instead of scanning dozens of trends," "instead of asking humans to stare at more dashboards").

Process plants run decades old instrument software with no modern interface, and screen scraping is a common answer there. Archimetis reads the historian directly and never touches the control network.

Sourcearchimetis.ai/refining Integrate your plant and Enterprise-Grade Security card 03, /chemicalsread 2026-09-13

Plant control systems are reached through a published industrial protocol, not through a browser or computer interface. The on premises controller talks to the DCS, laboratory and data historians over 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.

Interface changes pose no risk, because OPC-UA tags are a stable contract. The stated failure mode is instrument outage, and then the model turns itself off. The cloud Industrial AI platform is an application the customer's teams use, so a person operates that screen. The controller writes tags and does not drive a DCS operator console.

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