Nuclearn Request a meeting
Blog From the Nuclearn team

Your P&IDs already know the answer. Nuclearn Link makes them say it.

Somebody needs a pump out of service.

So an engineer pulls the system drawings. Finds the pump on a sheet. Starts walking outward along the process lines, valve by valve, deciding at each one whether it actually blocks flow. A gate valve stops the walk. A check valve does not. A power-operated valve closes, but seat leakage and fail position mean it is not a boundary on its own. Then the line runs off the edge of the drawing into a flag — continued on M-22EG02 — and the whole thing starts again on another sheet, which has to be found, printed, and matched up by eye.

Done carefully, that is a morning. Done on a plant where the P&IDs are twenty-year-old scans and the drawing index is a spreadsheet, it is longer. And it happens again next week, for a different component, starting from scratch.

What that work costs today

Idaho National Laboratory put numbers on a slice of it. In Scalable Methods to Automate Manual Work Management Activities Using Artificial Intelligence (INL/RPT-24-80159, August 2024), INL estimated that four drawing-dependent work-management steps — clearance order requests, and the reactivity, redundant-SSC and containment/heat-removal effect checks — run to roughly $440,000 per reactor unit per year at $200 an hour. That is INL's estimate of what those steps cost now, for online work only. It is not a figure for what any software saves, and INL notes its step times are expert estimates that vary by plant.

The obvious fix — redraw everything into intelligent CAD — has its own price. Conversion vendors publicly estimate eight to sixteen hours of drafting per sheet. For a plant holding a few hundred P&IDs, that is a project with a steering committee, and nobody gets an answer to this week's isolation question while it runs.

The drawings aren't the problem. The gaps between them are.

Each sheet is already a complete, correct description of its own piece of plant. What is missing is any machine-readable statement that this flag on sheet one is that flag on sheet two. A PDF has no opinion about what a valve connects to. A scan has even less.

That gap is why the work stays manual, and it is also why AI assistants have been useless at it. An assistant cannot follow a line across five PDFs. Ask it what isolates a pump and, at best, it reads you the drawing's title block.

What Nuclearn Link does

Link starts from the drawings you already hold — PDFs of any length, raster scans, image files, MicroStation DGN files, a zip of all of the above, and an optional index spreadsheet to name the sheets. No redraw.

It reads each sheet: the valves, pumps, tanks, instruments, and the pipe and signal lines between them, with pipes and instrument signal lines told apart. Then it joins the sheets at their off-page connectors, within one facility — links never cross between plants — and the whole drawing set becomes one connected map.

Now the morning's tracing is a question.

Pick a component, get its isolation boundary. Link walks the connected graph outward and returns a numbered list of the valves to close, following the boundary through linked connectors onto whatever other sheets the line reaches. In the worked example on our product page — using publicly available P&IDs from the NRC's ADAMS library — picking one valve returns nine valves to close across two sheets, with the second sheet one click away and the boundary still drawn on it when you get there.

It also returns the things the list alone would not tell you:

  • Drains and vents inside the boundary, for depressurising.
  • Relief paths the boundary reaches, to vent or account for.
  • Valves it does not credit — check valves, which block reverse flow only, and power-operated valves, which close but are not a clearance boundary.
  • Valves whose type it could not identify, counted as isolation points and flagged to confirm on the drawing.
  • Every open end, and every sheet-to-sheet link nobody has confirmed yet.

Follow a line off the sheet. Off-page connectors become links you can jump along and back. Each one is marked linked, suggested or missing. Where a match is uncertain, Link puts both sheets side by side and asks the only question that matters: does the line continue here?

Ask it from somewhere else. Link exposes a REST API and an MCP server, both using a Nuclearn platform API key, so AtomAssist, Claude, Copilot or your own agent can search the drawing library, find a component, follow a flag, and pull an isolation boundary across sheets. Same sign-in, same roles, same licence as the web app.

The part we are not going to bury

Link can misread a symbol. It can miss a pipe on a bad scan. It can propose a sheet-to-sheet link that is wrong.

We say so on purpose, because the alternative — a confident list with no provenance — is worse than no list at all. So the product is built to surface its own uncertainty rather than smooth it over:

  • Every export is explicitly marked as not clearance-ready.
  • The isolation view names each open end and counts the unconfirmed links, and says plainly when a boundary is not closed.
  • The MCP tools tell the assistant calling them that the result is not a clearance authority, so a chatbot answer carries the caveat too.
  • Electrical isolation is never on a P&ID. Link does not pretend otherwise.

And when it does get something wrong, you fix it on the drawing: mark a symbol wrong and it disappears, change its type, add a missing symbol, draw a missing line, set a valve's normal position. Every correction is recorded with who made it and when — and it carries over the next time Link reads the drawings, including after a product update. The corrections your people make are an asset that accumulates, not work you redo.

That human step is the product working as designed. Link's job is to turn a blank page into a checked list. The clearance writer's judgement is still the thing that writes the clearance.

What it is not

Link is not a CAD redraw, and it is not a clearance or tagout system. It is the connected map those jobs start from. Check its boundaries against the drawings before relying on them — the product is built to make that check fast, which is the point.

Each plant gets its own single-tenant instance, and people sign in with their existing Nuclearn account, so the platform's SSO, roles and licence apply. Link keeps no users or passwords of its own. Nuclearn is ISO/IEC 27001:2022 certified and has built AI software for nuclear plants since 2021; Nuclearn products serve over 70 facilities across North America.

See it on a real drawing set

The fastest way to judge this is to watch a boundary cross a sheet. Book a demo at link.nuclearn.ai and we will scope an isolation boundary across linked sheets on publicly available plant P&IDs, show connected components and linked drawings, and let an AI assistant answer from them. Bring your IT team's questions on deployment and sign-in — those get answered in the same call.

Then bring us one of your own ugly scans and see what it does.

Subscribe

New posts by email.

  • Each new post from the Nuclearn blog, as it's published or in a digest
  • Press releases and product news

You can unsubscribe at any time from the link in each email. See our Privacy Policy. Want to talk to the team instead? Request a meeting.

Subscribe to the blog

Your email is all we need.