02:14 on a Tuesday. A ground robot has just finished the turbine building operator rounds route, unattended, in the dark. It read 43 points. 41 are nominal. Two are not: a thermal anomaly on a main steam line that has grown 18 degrees C across four passes while still sitting under its alarm limit, and a lube oil reservoir gauge that is below the low limit. GridCORTEX compared every one of the 43 readings against the full history for that exact point, raised two exceptions, and drafted one work request with the trend and the thermal image attached. The operator who used to walk this route at 2 AM reviews two exceptions in four minutes instead of walking for ninety. Then watch the same intelligence layer take a radiological survey, a confined space inspection with no scaffolding, and an underwater ROV dive, and drop all of it into one findings queue.
| Without | With GridCORTEX |
|---|
| Without | With GridCORTEX |
|---|
The stage is a top down floor plan of a fictional turbine building. The green line is operator rounds route 3, the same route a person walks on the back shift today. The 43 small markers on it are the readable points: local gauges, thermal targets, gas readings, level glasses, fire equipment seals and visual checks. They start dim, meaning not yet read this pass. A robot marker walks the route from the dock in the southwest corner, and each point lights as it is read: green for nominal, amber for a reading outside a limit, red for a reading that is inside its limit but moving the wrong way against its own history. The panel on the right shows the point being read right now: its tag, what it is, the value, the limit, and five bars, which are the last four passes plus this one.
The route starts at 01:31 and finishes at 02:14. At point 23, on the main steam corridor at elevation 3, column line H, the robot reads an insulation joint surface temperature of 91 degrees C against a 95 degree limit. That passes. The last four passes at that same point read 73, 78, 83 and 87, so the joint has climbed 18 degrees C across four passes and is heading for the limit on a schedule you can draw a line through. That is exception one, and it is the one that becomes a draft work request. At point 29, in the lube oil room, the turbine lube oil reservoir reads 62 percent against a 65 percent low limit, with the prior four passes at 71, 69, 67 and 65. That is exception two. Point 30, four steps away, is a floor drain sump check that came back "trace sheen" for the second pass running. That one is nominal on its own, so it is not a third exception, but it is attached to the oil exception as supporting evidence because the same pass saw both.
At 02:14 the route closes. GridCORTEX packages what it has: 41 nominal readings written back to the rounds record so the round is documented as performed, two exceptions, and one draft work request for the steam line with the thermal image and the five pass trend attached. At 02:18, four minutes later, the operations supervisor on shift opens the rounds board, looks at two cards, and approves. Nobody walked the building. Then the simulation keeps going, because rounds are only the first route this layer runs. Wednesday it flies a radiological survey of a controlled area and returns a three dimensional dose map. Thursday it puts a collision tolerant drone inside condenser waterbox 1B with no scaffolding erected and no confined space entry made. Friday an ROV surveys the intake structure and the traveling screens instead of a diver. Every one of those findings lands in the same queue as the two from Tuesday night, deduplicated against each other and against what is already open in the CMMS. That single queue is the product. The robots are the sensors.
You may already own the robot. Plenty of utilities do, and some own whole inspection fleets and fly them commercially. What you get from it is a folder. The robot walks the route and produces 43 readings, a few hundred images, and a completion timestamp, and then a person has to open all of it and decide what matters. The steam line reads 91 against a 95 limit and passes, because nothing on the sheet compares it to the last four passes, and the last four passes are in a binder, a scanned PDF, or a rounds system nobody queries that way. The oil level is low, gets written up, gets topped off, and the sheen in the sump three points later never gets connected to it. Meanwhile the radiological survey lives in the RP group's system, the waterbox inspection lives in a contractor's PDF report, and the ROV video lives on a hard drive in a cabinet. Four inspection programs, four queues, no comparison, no dedup. That is the state most robotics programs reach and then stall in.
The layer above the robot does four things a robot does not do. It compares: every reading against the full history for that exact point, and against the instrumented historian tag when one exists nearby. It ranks: two exceptions out of 43 readings, ordered by consequence, with everything else filed as evidence rather than as work. It drafts: one work request in your CMMS in pending status, with the trend, the image and the exact location attached, waiting for a named approver. And it consolidates: rounds, radiological survey, confined space inspection and ROV findings arrive in one queue, deduplicated against each other and against your open condition reports, so a wall you have already written up does not get written up three more times by three different robots. GridCORTEX is vendor neutral by design: it ingests from whatever fleet or contractor you use and it never operates plant equipment.
These are measurements a pilot produces, not results we are claiming. Each row names the number, how it is measured, what it is compared against, and who at your plant has to agree the number is real. The demo above is synthetic; none of these cells contains a value, because the value is yours to produce.
| What the pilot measures | How it is measured | Compared against | Who signs the number |
|---|---|---|---|
| Reading accuracywhether the robot read the gauge correctly | Robot reading versus the human reading taken on the same point during the parallel run | The operator walking the same route the same shift | The operations supervisor who owns route 3 |
| Instrumented agreementthe objective check where one exists | Robot reading versus the historian tag value at the same timestamp, for points that are both walked and instrumented | The plant historian, which nobody disputes | The historian data owner |
| Catch ratedid it find what a person finds | Count of findings the human raised that the robot also raised, over the parallel run | The human's own rounds sheets for the same passes | The operations supervisor |
| Additional findsdid it find anything a person does not | Count of trend based exceptions raised on points that were inside limits every pass | The prior rounds record, which is why we ask for two to three years of it | The system engineer for the affected system |
| False exception ratehow much noise the operator has to sort through | Exceptions the reviewer marks "not real" divided by all exceptions raised | The reviewer's own judgment, logged per exception | The shift operator doing the review |
| Route completion ratehow often the robot actually finishes | Completed passes divided by scheduled passes, with a reason code on every miss | The schedule you set, not an availability target we set | The operations supervisor |
| Operator time on the routethe hours number the calculator needs | Stopwatch on the human walk before, and exception review time logged per pass after | Your own before measurement, taken during the parallel run | The operations supervisor and your finance partner |
| Handling overheadthe cost side nobody puts in the deck | Hours per week logged for charging, recovery, route edits, and dealing with a blocked route | Nothing. This is a new cost and it is measured as one | Whoever ends up owning the robot day to day |
| Dose on controlled area passesonly if the route enters a controlled area | Person-rem that would have been accrued by the human pass, from your own dosimetry records for that route | Your ALARA plan and your historical dose for the same work | Radiation protection |
| Confined space entries and scaffold hours avoidedthe units your safety group already tracks | Count of permits not written and scaffold build hours not spent for inspections the robot covered | Last outage's permit log and scaffold invoices | Your safety lead and the outage manager |
| Duplicate findings suppressedwhether the one queue idea actually works | Count of incoming findings matched to an already open condition report or work order instead of raised again | Your CMMS backlog on the day the finding arrives | The CMMS administrator |
| Time from finding to approved work requestwhether the queue moves | Timestamp of pass completion to timestamp of approval, per finding | Your current time from inspection to work request, measured before the pilot | The work management lead |
Type your own numbers into the boxes. Everything below them recalculates as you type, and every result prints the arithmetic with your numbers substituted in, so you can check it on paper or hand it to your finance partner without taking anyone's word for it. The values sitting in the boxes right now are placeholders, chosen as examples, and they are not claims about your plant or anyone else's. Overwrite them.
Rounds put a person next to hot, pressurized, rotating and energized equipment on a schedule, at night, alone, in weather. Most rounds are uneventful, which is exactly why the exposure is easy to underweight. Sending the machine to the routine points means the human is only in that space when there is a reason to be there.
Counted in units you already track:
The rounds hours themselves come back to your operations staff, and the transcription and filing time comes back on top of that. Your operator does not stop working, they stop walking, and the time goes to the exceptions and to the work that actually needs a licensed person.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Rounds labor | rounds hours avoided x your loaded operator rate, with the night and weekend portion at your actual shift differential |
| Transcription and records | transcription hours avoided x your loaded rate, plus whatever your records group spends filing and retrieving rounds sheets |
| Emergent to planned conversion | your own average cost of an emergent repair minus your average cost of the same repair planned, multiplied by the number of findings you believe earlier detection converted. You set that number, we do not |
| Insulation and steam loss | your steam cost per hour x the hours a leak or missing insulation would have run undetected between calendar inspections. Your plant already knows what a steam leak costs per hour |
| Dose reduction value | person-rem avoided x your own internal dollar per person-rem, which most nuclear operators already carry for ALARA decisions |
You pay for the robot or the inspection service, for the edge compute and the site network coverage on the route, for the GridCORTEX layer that reads and trends the results, for the CMMS and historian integration, and for real staff time during the parallel run. You will also pay in unglamorous ways: charging docks, door and stair access, and someone who owns the robot when it gets stuck. Plants that skip that last item are the ones where the program stalls.
Payback is driven by rounds labor first and by emergent to planned conversion second. Build the case on the labor, which you can audit from your own rounds schedule, and treat the avoided failures as upside until your first year of actuals says otherwise.
The person who makes the dose map is the person who takes dose making it, and they take it in the areas that are hottest by definition. Sending a machine to build the map removes that exposure entirely. The second and larger effect is downstream: every crew that later works in that space is routed by a model that sees the vertical structure of the field, so they take less dose doing the actual job.
Counted in units you already track:
Health physics survey hours come back, ALARA planning hours come back, and the outage planner stops rebuilding dose estimates by hand for every job in the area.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Survey labor | survey hours avoided x your loaded health physics technician rate |
| Dose value | person-rem avoided x your own internal dollar per person-rem. Most nuclear operators already carry this figure for ALARA decisions, and if you do not, the case should be built without it rather than with a number we invented |
| Planning labor | ALARA planning hours avoided x your loaded coordinator rate |
| Outage critical path | hours removed from a critical path job x your own cost per outage hour. Only count this where the job is genuinely on critical path, which your scheduler can tell you |
| Dose limit headroom | not a dollar figure, a constraint. If dose budget is what stops you from doing a job this cycle, freeing person-rem is what lets the job happen, and the value is the deferred work you can now execute |
You pay for the survey platform or the survey service, for the instrument and its calibration program, for the GridCORTEX modeling layer, for integration into your RWP and dosimetry systems, and for a genuinely serious instrument validation phase up front. That validation phase is not optional and it is not fast. If your radiation protection group does not trust the readings, the model has no value at any price.
Payback is dominated by dose value and by outage critical path hours, in that order, and both are numbers you already track. Survey labor alone will usually not carry the case, so do not build it on labor.
Confined space entry is among the highest consequence routine activities in any plant, and the industry knows it: the fatalities are rare and catastrophic and a meaningful share of them are would-be rescuers. Scaffolding adds falls from height and dropped object exposure on top. Inspecting without entering removes the entry, the standby, the rescue scenario, and the scaffold, and leaves them for the jobs where a person genuinely has to be in there.
Counted in units you already track:
The hours that come back are mostly not the inspector's. They are the scaffolders', the attendants', the rescue standby's, and the operations staff who write the isolation. That is the part plants consistently underestimate when they price this.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Scaffolding | scaffold hours avoided x your loaded scaffolding rate, or your contracted scaffold cost per erection if you buy it that way. This is usually the largest single line and most plants can pull it straight from job costing |
| Standby and permit labor | attendant, rescue standby and permit writing hours avoided x your loaded rates |
| Isolation and operations labor | isolation and lockout hours avoided x your loaded operations rate |
| Outage duration | outage hours removed from critical path x your own cost per outage hour, counted only where your scheduler confirms the inspection was on critical path |
| Dose value | person-rem avoided by scaffolders and inspectors x your internal dollar per person-rem, where the space is in a controlled area |
You pay for the inspection platform or the inspection service, for the GridCORTEX layer that locates, measures and trends findings, for integration into your asset and inspection data systems, and for a serious engineering acceptance phase where your inspection authority decides which inspection types the robotic method can satisfy. That acceptance work is the real cost and the real gate. Skipping it produces a fleet of drones and an inspection program that still writes permits.
Payback is dominated by scaffolding and standby labor, both of which you can pull from job costing, and by outage critical path hours where they apply. Build the case on those. Everything else is upside.
Commercial diving is a high consequence activity in any industry, and diving in a nuclear plant adds radiological exposure to everyone in and around the water. Substituting a submersible removes the diver, the standby rescue diver, and their dose entirely for the scope it can cover. It also removes the pressure to keep a diver in the water longer than planned when a search runs long, which is the situation where dive incidents actually happen.
Counted in units you already track:
Dive team hours, dive support hours and radiation protection coverage hours come back, and the search dives, which are the ones that blow out the schedule, become vehicle time instead of human time.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Dive services | dives substituted x your contracted dive cost. The Department of Energy study across seven utilities put the combined saving at roughly ten thousand dollars per dive including dose reduction, which is a published figure you can sanity check against your own contract rather than a number we generated |
| Dose value | dive team person-rem avoided x your internal dollar per person-rem |
| Support and coverage labor | supervisor, surface support and radiation protection coverage hours avoided x your loaded rates |
| Outage duration | critical path hours removed x your cost per outage hour, counted only where the dive was on critical path, which for foreign object searches it very often is |
| Intake availability | your cost per hour of derate or reduced circulating water flow x hours of derate avoided by catching intake debris accumulation between cycles rather than at the next scheduled cleaning |
You pay for the submersible or the ROV service, for radiological controls and decontamination on the vehicle, for foreign material exclusion qualification of the vehicle itself, which for pool work is the gating item, for the GridCORTEX layer that reads and trends the surveys, and for integration into your asset and FME systems. Expect the qualification work for pool use to take longer than the technology work. That is appropriate and you should plan for it rather than fight it.
Payback here is more direct than any other physical AI use case in this set, because the dive substitution count times your own contracted dive rate is a single line of arithmetic against a schedule you already have. Start there, and treat dose value and critical path as the confirming case.
This use case does not directly remove a person from a hazard. It does something else that matters just as much: it makes sure the findings that came from removing people from hazards actually reach the work stream. A defect found by a drone and never turned into a work order delivered no safety benefit at all, and the engineer whose inbox it died in is not the problem.
Counted in units you already track:
The hours that come back are engineering review hours and planner re-keying hours, and they come back from the most experienced people you have, which is why they are the expensive hours.
The numbers we need from you to run that formula:
| Cost driver | How it is calculated, from a rate you supply |
|---|---|
| Engineering review labor | review hours avoided x your loaded inspection engineer rate |
| Duplicate work eliminated | duplicate work orders avoided x your fully loaded average work order cost, including the craft hours and the mobilization, not just the planning |
| Re-keying and administration | re-keying hours avoided x your loaded planner rate |
| Program rationalization | the inspection spend you redirect once you can see which asset classes and which platforms never produce an actionable finding. This is often the largest number and nobody can calculate it today because nobody can see it |
| Missed finding avoidance | your own cost of a failure that a received but unread finding would have prevented, x the number you believe applies. You set that number. We will not |
You pay for the GridCORTEX layer and its integrations, and for engineering time to label past findings so the ranking is yours rather than generic. You do not pay for new robots, and that is the point. If your findings are arriving as contractor PDFs, budget for extraction work per format, which is unglamorous and real. Where a finding class is rare enough that there is not enough real imagery to train a detector, NVIDIA Omniverse Replicator can generate physically accurate synthetic images to fill the gap. Exelon did exactly that with Deloitte for utility pole cross-arm defects because real labeled defect data was too scarce.
Payback is engineering review hours plus eliminated duplicate work, both of which you can measure inside one quarter. Program rationalization is usually the bigger number but it takes a full cycle of data before you can defend it, so do not put it in the first business case.
What is this, exactly? It is AI software: intelligent agents and models built and delivered by SoftServe, running on NVIDIA accelerated computing. It is not a hardware appliance and it does not replace the systems you run today. It deploys in your own cloud or on your premises, connects read-only to your existing systems, and recommends; your people approve every action, starting in shadow mode until it earns trust.
An autonomous rounds program. The robot follows the same route your operator follows, on the schedule you set, and GridCORTEX reads every gauge, image and sensor reading it brings back, compares it to the history for that exact point, and raises a finding only when something is outside your own limits or is trending toward them. Your operator stops walking the route and starts reviewing exceptions. The demo above uses synthetic data; everything below describes what the real deployment needs from your organization.
| Your system | Typical products | How we connect |
|---|---|---|
| Robot or drone fleet platform | Boston Dynamics Orbit, ANYbotics ANYmal, Percepto AIM, Flyability Inspector, or your inspection contractor's platform | vendor API or file drop, fleet agnostic |
| Computerized Maintenance Management System (CMMS) | Maximo, SAP PM, Passport, Hexagon EAM | write API, notification or work request in draft status |
| Plant historian | AVEVA PI System, GE Proficy, AspenTech eDNA | read-only tag subscription so robot readings are checked against instrumented values |
| Operator rounds or logging system | eSOMS, Maximo rounds, paper or tablet rounds sheets | read for route definition, write for completed rounds record |
| Fire protection and safety equipment register | CMMS module or standalone register | read-only, to match imaged equipment to its inspection record |
The robot and its edge compute live on plant property. Inference runs at the edge on NVIDIA Jetson so imagery never has to leave the site. GridCORTEX runs on-premises or in your own cloud account, connected read-only to the historian and with a write connection to the CMMS that creates draft work requests only. Nothing connects to safety related systems and nothing issues a control action. The robot has no ability to operate plant equipment.
The Approve button you just clicked in the demo above is the real workflow. This is what it looks like on the screen of the operations supervisor and the shift operator who used to walk the route:
Approve creates a draft work request in your CMMS through its API, in pending status, attributed to the approver. It does not create a work order directly, it does not assign a craft, and it does not touch the schedule. If the reading is one your rounds system requires to be logged, the value and its timestamp are written to the rounds record. Nothing is written to the historian, to the plant control system, or to any safety related system, ever.
The route and its cadence are set once by your supervisor. After that the operator gives input in two ways: marking a flagged exception as real or not real, which is how the detector improves, and adding a one-off route when something needs eyes on it, for example after a maintenance activity or before a walkdown.
Findings are as fresh as the last completed pass and every card shows the pass timestamp. A route walked at 02:14 is labeled 02:14, not "current". Comparisons against instrumented historian values use live tags. If the robot is down or a route is blocked, the console shows the route as stale with the reason, because a rounds program that quietly stops is worse than no rounds program.
A rounds board in the GridCORTEX console showing every route, its last completion, and its open exceptions. A single morning digest to the operations supervisor. A phone push only for a reading outside a limit you have marked as immediate. The console runs in a browser beside your existing screens on day one; embedding into your own systems is a roadmap step once the read-only phase has earned trust. Approve, Modify, and Decline are all captured in an audit trail your compliance team can pull, and GridCORTEX never blocks or overrides anything in the systems you run today.
The fair question from any plant manager or ops director: "We already have a rounds program, a historian, a CMMS, and in some cases we already own robots. What is actually new here?" Here is the honest answer, and it is why most inspection robotics programs stall after the first successful demo.
When someone asks "what did it actually calculate?" this is the list. In the simulation these factors drive the storyline; in a pilot they are computed from your rounds sheets, your reading history, your limits and your CMMS backlog.
Presenter's one-liner: "The robot walked 43 points at two in the morning and 41 of them were fine. The two that were not are the whole product: one was outside a limit, and one was inside its limit and climbing 18 degrees across four passes, which you can only see if something remembers the other four passes. Four minutes of a supervisor's attention replaced ninety minutes of walking, and the same queue now holds the radiological survey, the waterbox inspection and the ROV dive."