AI PPAP Generation: How AI Builds a PPAP Package and Keeps Engineers in Control

Ravi Kumar

Welcome to Deep Dive 360, where we look at where quality, manufacturing and technology meet. I'm Ravi Kumar, CTO of Aekam AI. I've spent over 25 years building enterprise software, which means I usually show up with the technology and then ask the shop floor whether it actually works. And today I've brought someone who will tell me.

Abhi Ram

Happy to be that person. I'm Abhi Ram, Chief Product Development Officer at Aekam AI. I've spent more than 15 years in manufacturing and quality, mostly on the problems that never show up in a slide deck. Certificates that don't match the spec. Inspections that miss things. And paperwork that slows good parts down.

Ravi Kumar

And today the paperwork is the topic. Pee pap. The Production Part Approval Process. Specifically, can AI generate a pee pap, how does it work, and how do you stay in control of it? Quick context for anyone new: Aekam AI builds AI products for manufacturing quality, and one of them is Aekam AI pee pap Studio, which generates pee pap packages. So Abhi, let's start with the real problem.

Abhi Ram

Here's what actually happens on the floor. It's Thursday, 4:30 in the afternoon. An email arrives from the customer. Pee pap for the new part is due Monday. Level 3. Their workbook, not yours. And they want C P K on every special characteristic.

Ravi Kumar

I can already feel the weekend disappearing.

Abhi Ram

It is gone. You open the drawing and number every dimension by hand. You copy those numbers into the dimensional report. Then the control plan. Then the PFMEA. You chase the inspector for the handwritten sheets and type every reading into Excel. On Sunday night you're pasting it all into the customer's template, cell by cell. And here's the thing I noticed over fifteen years. The real engineering decisions take an hour or two. In my experience, about seventy percent of the time is repeatable clerical work.

Ravi Kumar

So from my side, that's not a quality problem. That's a data problem. The same characteristic gets typed into five documents by hand, and every retype is a chance to get it wrong.

Abhi Ram

Exactly. Balloon thirteen on the drawing shows up in the dimensional report, the control plan, the P F M E A and the capability study. Same number, five times.

Ravi Kumar

Which is the kind of problem software should be very good at. So walk me through it. When people hear "AI writes the P F M E A", what do they picture?

Abhi Ram

A chatbot inventing failure modes. And honestly, they should be worried about that. That's not how we built it. Think of it like this. A general AI model is a cook who has read every cookbook in the world. Your plant has its own recipe book. Your piercing operation runs on a specific press. Your patrol check is five pieces every two hours. You had a burr problem on a similar part last year. The general cook doesn't know any of that.

Ravi Kumar

So the recipe book is what we call the knowledge bank.

Abhi Ram

Right. Your operations with their control methods and reaction plans. Your failure modes. Your machines. Your gauges, with calibration dates and gauge R and R results. Your sampling rules. Your past pee paps. And your customers' templates. Most plants already have all of this. It's just spread across spreadsheets, old control plans, and the head of one senior engineer.

Ravi Kumar

And technically, that's the important design choice. The AI isn't answering from general knowledge. It's grounded in the plant's own data, and each customer's knowledge bank is their own. Every time an engineer approves a change, say a new failure mode, the system offers to save it back. But it only goes in after a person approves it. So the bank gets better with every package, and nothing changes behind anyone's back.

Abhi Ram

Which matters, because that's what makes the next package faster than the last one.

Ravi Kumar

Okay. Step one in practice is the drawing. And drawings are messy. Overlapping callouts, crooked scans, someone's handwritten note in the corner. How does it cope?

Abhi Ram

The AI reads the whole sheet, the way an engineer scans a print before ballooning it. Every dimension, tolerance, G D and T callout and note becomes a numbered characteristic, and a balloon goes on the print next to it. It also picks up the critical and significant characteristics and cross-checks them against the customer-specific requirements, so you don't miss a critical feature because the symbol was faint.

Ravi Kumar

And when it isn't sure?

Abhi Ram

Every characteristic carries a confidence score. Anything below the threshold gets a dashed balloon and goes into a review queue. The engineer checks it against the print, fixes it if needed, and approves the list. And let's go back to the basics here: that balloon list is the spine of the whole pee pap. Everything downstream hangs off those balloon numbers.

Ravi Kumar

Which means revisions get cheaper too.

Abhi Ram

Much cheaper. When the customer sends Rev C, the system compares it to the approved Rev B and shows only what changed. You update four characteristics, not ten documents.

Ravi Kumar

Now, every plant I visit still runs on clipboards. Is that a problem?

Abhi Ram

Not at all, and we didn't want to change that. The inspector measures and writes exactly as today. They take a photo of the sheet, or scan it. The AI reads each value, puts it in the right row, and checks it against the tolerance at the same time. And it doesn't guess. If a digit could be a one or a seven, it's highlighted on the photo and in the table, and the engineer decides. Same for gauge R and R registers and S P C sheets. Then the system does the math: percent G R R, number of distinct categories, C P, C P K, P P and P P K, written into your own templates.

Ravi Kumar

Alright, here's where it gets interesting. The control plan and the P F M E A. That's where the engineering judgement lives. How does the AI create those without making things up?

Abhi Ram

It assembles facts. It doesn't write from scratch. The process flow is suggested from your closest past parts, with a reason for every step. Then every control plan row answers four questions. Which operation? That comes from the process flow. Which characteristic? From the drawing. How is it measured, how often, and what's the reaction plan? From your own rules in the knowledge bank.

Ravi Kumar

And from a data point of view, every cell keeps its lineage.

Abhi Ram

Yes, and you can see it. Each cell is colour-coded by its source: the drawing, the process flow, the knowledge bank, a handwritten reading, or an AI inference. Anything inferred or low confidence is amber. The engineer clicks "next to review" and goes through only those cells. Then the P F M E A is drafted from your failure-mode library. Severity goes up wherever a critical characteristic is produced. Detection is matched to the measurement method the control plan actually uses, so the two documents can't contradict each other. And you pick the method: the classic R P N, or the A I A G, V D A action priority.

Ravi Kumar

I have to ask the question every quality director asks. What happens when the AI gets it wrong?

Abhi Ram

It will, sometimes. I'd be suspicious of anyone who says theirs never does. Manual pee paps are full of mistakes too. The real danger is a clean-looking spreadsheet that hides which values were guessed. So we built it on four rules. One: nothing is invented. Every value traces to a document, the knowledge bank or a real reading. Two: uncertainty is shown, not hidden. Three: every element is approved separately, and the next one waits. Four: the engineer signs the P S W, not the software.

Ravi Kumar

And when someone fixes something?

Abhi Ram

You fix it where it happened, and the change flows downstream. Correct a tolerance on one balloon, and the dimensional report, the control plan and the capability rows that use it are flagged for update. Every change sits in the package history. That's exactly what an I A T F auditor wants to see.

Ravi Kumar

There's one more piece I think people underestimate. The system doesn't just build the package, it checks it while it's being built.

Abhi Ram

That's my favourite part. Think about how a pee pap really fails. It's rarely a missing document. It's a reading out of tolerance hiding in a two-hundred-row table. A C P K below what the customer demanded. A gauge whose R and R study expired. A mill certificate that doesn't actually meet the grade on the purchase order. All of that can be checked. Usually nobody has the time.

Ravi Kumar

So the customer's documents become the source of truth.

Abhi Ram

Right. The drawing, the P O, the customer-specific requirements and the specs. An out-of-spec reading is flagged the moment it's read. If the customer wants a C P K of one point six seven and a characteristic comes in lower, approval stops until there's an action plan. Material test reports are checked property by property against the drawing notes, the customer spec, the A S T M or internal standard, and the P O grade. You find the problem on your floor, not in the customer's rejection email.

Ravi Kumar

Last objection, and I hear it every time. "Our process is different."

Abhi Ram

And it usually is. That's why the workflow is defined by your data. Your operation numbers. Your reaction plans. Your gauge rules. Your F M E A method. Even outsourced steps like plating. And the customer's template: you upload their Excel workbook once, the system maps it, and every future package for that customer goes straight in. No copy and paste on Sunday night.

Ravi Kumar

So, the question my finance team would ask. Is it worth it?

Abhi Ram

Let me keep it simple. Count your pee paps per year. Multiply by the hours you save on each one. In our time model, a typical Level 3 package goes from about thirty-five engineer hours to about six. But don't take our model for it. Run it on a part you've already submitted, time it, and compare. And beyond the hours, you get fewer rejections, faster revisions, and your pee pap know-how stops living in one person's head.

Ravi Kumar

If you had to give a quality director one sentence to take away?

Abhi Ram

AI should do the copying, and your engineers should do the engineering.

Ravi Kumar

I like that. If you want to see it on your own part, send us a drawing, ideally one you've already done a pee pap for. We'll show you the ballooned print, the control plan and the P F M E A generated from it, with every source and every flag visible. You can book a demo at aekam dot A I. Abhi, thank you.

Abhi Ram

Thank you, Ravi. And to everyone listening: keep your weekends.

Ravi Kumar

That's it for this episode of Deep Dive 360. See you next time.