The first question a fabrication shop owner asked me about RevE wasn't about accuracy, and it wasn't about cost. It was this:
"Where does my customer's drawing go?"
It's exactly the right question. A fabrication drawing is a contract document. It carries the end client's name, the project, the plant, the people who drew, checked and approved it, document numbers, order references, sometimes a logo and a signature. Many shops sign confidentiality clauses that say, in plain words, that the drawing doesn't leave their hands.
And an AI drawing reader, by its nature, sends drawings to a model in the cloud.
This is how we resolved that tension without a private data centre, without a local AI model and without hiding anything from the user. It comes down to one sentence we agreed on early and never broke: the AI may read the engineering of a drawing, but not whose drawing it is.
First, be honest about what you're sending
It's comforting to believe an AI system sends "only bits" of a drawing: a crop here, a table there. When we actually added it up, the truth was different. One full reading sends an overview of every sheet, a close-up of every block (twice, because we read everything twice), and all the text that was read. Put together, that's the complete drawing, in pieces.
I think many AI projects skip this step. Do it. It changes the conversation from "we only send what's needed" to "we send everything, so what exactly must we protect?"
Two kinds of secret
Once we accepted that the model sees the whole sheet, we looked at what on that sheet is actually secret. There are two very different kinds.
Identity: who the drawing belongs to and where it's going. That means the customer, the end client, the project, the plant, people's names and initials, approval dates, document and order numbers, equipment tags, logos and signatures. This is what confidentiality clauses mostly protect. Crucially, it's also information the AI doesn't need. To read a design pressure, it doesn't need to know whose plant the exchanger is going to.
Design know-how: sizes, materials, pressures, temperatures, thicknesses and nozzle ratings. This is the engineering, and the AI must read it, or the whole system is pointless.
That split gave us a strategy. Hide identity completely, before anything leaves the laptop. Send the engineering, with the protections a commercial AI service offers, and be completely open about it.
Why not just run the AI locally?
It's the answer everyone expects, and I wanted it to work. But this customer will run RevE on a Windows laptop with 4 to 8 GB of memory, and a vision model good enough to read engineering drawings simply doesn't run there. For the most sensitive cases, the realistic alternative is an AI service in the customer's own cloud account, and we've kept that as a future option. Pretending otherwise would have meant a demo on my Mac and disappointment on theirs.
How the protection works
Everything in this section happens on the laptop. No AI is involved.
Find the identity. RevE looks next to familiar captions (CUSTOMER, PROJECT, DRAWN, CHECKED, APPROVED, DATE and so on), down the columns of revision tables, and for patterns like dates, e-mail addresses and company names. It adds names from the user's own list, flags small pictures that look like logos or signatures, and lets the user cover any area by hand. On our five sample drawings it found between 7 and 124 identity spots per drawing. Most of the big numbers came from revision tables, with a date and initials in every row. Revision tables turn out to be an identity goldmine that most redaction approaches ignore.
Make a protected copy that's really protected. There's a classic redaction failure: someone draws a black box over text in a PDF, and the text is still there underneath, searchable. RevE doesn't draw over the text. It removes the letters under each spot, then places a grey cover labelled with a code: CUSTOMER-1, DATE-6, PERSON-3. It also empties the file's hidden information, where names often lurk.
Check for leaks before anyone approves. We didn't trust our own redaction, so RevE checks it. It confirms that no hidden text is left, re-reads every covered spot on the laptop to make sure nothing is legible, and confirms the file information is empty. A leak blocks approval. To prove the check works, we planted a leak on purpose, and it was caught.
A person approves. The user sees a screen called "What the AI will see", with the protected sheet, every covered spot listed by kind, and buttons to hide a word everywhere or cover an area. When they approve, the approval is tied to that exact protected copy. Any change cancels it, and without a current approval the AI run simply won't start.
Protecting every request, not just the pictures
Pictures are only half of it. Text travels too: chat questions, and text read in earlier steps. So the protection sits in the one function every request passes through. Just before sending, names are swapped for their codes. When replies come back, the codes are swapped back to real names, on the laptop. The table linking codes to names never leaves the machine, and requests are sent with storage switched off on the provider's side.
The effect is simple. The user sees real names everywhere in the app. The model only ever sees CUSTOMER-1 and PERSON-3. In our test, 41 of 41 hidden values came back correctly in the final results, and no code was left behind in any file.
Every request is also recorded in an outbox the user can open and export: which drawing, which step, what was sent and how many names were hidden. Security people don't just want protection. They want evidence.
The newest step: a chat that answers on the laptop first
RevE has a chat: "What's the shell thickness?", "List the nozzles", "What's the design pressure on the shell side?" Originally, every question went to the online model with the drawing's reading as context.
Then the user asked a sharp question: "When I ask the chat something, are we sending our drawing data to the AI every time?" The honest answer was yes, protected but still sent.
So we rebuilt the chat to answer on the laptop first. Most chat questions are really lookups, and simple keyword search over what RevE has already read can answer them in about a millisecond, at no cost and with nothing sent. It even copes with shop-floor typing: "List of nobles" finds the nozzles, and "shell thk" finds the shell thickness. Each answer is marked "on this computer" and shows exactly where the value came from. A question that has nothing to do with the drawing ("What's the capital of India?") is politely turned away, and nothing is sent.
Only questions that need reasoning, like "what if" or "why", go online. Even then only after the user presses a button, and only the 30 best-matching pieces of the drawing are sent, a few kilobytes instead of the whole reading. A "show what will be sent" link displays the exact text first, with names already replaced by codes.
What real drawings added to the job
Real drawings kept widening what "identity" means, and the protection had to follow.
- Names in a second language. Some drawings repeat a customer's name in Hindi beside the English. Hiding only the English one would leave the secret in plain sight. RevE now converts older Hindi fonts to proper text, finds those Hindi twins and hides them too (see Part 7).
- Initials that wander. A person's initials in a revision table are hidden. But a short string of letters like that can appear by chance inside an ordinary word elsewhere on the sheet. The leak check now accepts initials only inside their own covered spot, and flags them anywhere else.
- Drawings with many sheets. On large drawings, the approval screen now has a sheet picker and zoom, so the person approving can actually inspect every covered spot rather than scrolling past a tiny preview.
- Names put back safely. When replies come back, codes are turned back into real names by structure, not by a blind search-and-replace. An early version once tripped over a code that happened to match an ordinary word on the drawing. It's the kind of bug you only meet on real data.
The honest limits
We wrote down what this doesn't do, because users trust a system more when it tells them where to be careful.
- The identity finder is rule-based. A name with no caption beside it, in an unusual place, may be missed. That's why a person reviews the protected copy before approving.
- The app's log-in protects the app, not the files on disk. For that, use the computer's own disk encryption.
- The engineering itself still goes to a cloud model, under the provider's protections. For drawings where even the engineering is too sensitive, the right answer is a model in the customer's own cloud account, and we say so plainly.
How to explain it to your customer's customer
A shop has to answer to its own clients, often big companies with security teams. The technical design only helps if it can be explained in a few sentences on a call. This is the paragraph we settled on:
"The drawing never leaves our office with your name on it. Before any AI service sees it, our software removes your company, project, plant, people and document numbers, checks that nothing is left, and a person approves the protected copy. The AI reads only the engineering, under the provider's no-storage terms. Simple questions are answered on our own computer. We keep a record of every request, and we can share it with you."
That paragraph matters as much as the code. It's honest about what's sent, it names the protections in order, and it offers evidence instead of asking for trust.
The shop owner asked a simple question: where does my drawing go? We can now answer it precisely, and that's what turned a promising prototype into something a shop can use with its customers' drawings.
Next, in Part 10, the last of this series: the demo works on my machine, but the product has to work on theirs.
#DataPrivacy #IndustrialAI #Cybersecurity #Confidentiality #ResponsibleAI #Manufacturing #EngineeringDrawings #DataProtection #TrustworthyAI #SandurTech
Disclaimer
This series introduces SandurTech RevE, a product in development for AI-assisted engineering drawing analysis and generation. Examples reflect our current capabilities and ongoing validation. Outputs should be reviewed by qualified engineers and verified against approved drawings, standards, and project requirements before use.
