Automated Leather Inspection Machine, Kasra Zamanian

Setup

My thesis on visual defect detection for luxury leather impressed the client enough that they wanted a real product. I joined Corporate Hangar as the AI lead for their AI projects to build it, and we partnered with a robotics company to prototype a machine using linear cameras instead of the matrix cameras from the thesis. The complexity exploded immediately. It was no longer dots on one leather: there were tens of leather types, each with its own defect vocabulary, so different that transfer learning from one to another often hurt more than it helped.

Why it was hard

Two walls at once. The first was annotation: we needed expert operators to label, but operators did not want to annotate on images, they were used to touching and flexing the hide to find defects. And because this is luxury goods, a defect is contextual: the same flaw is fine in an offcut but unacceptable where a handbag panel will be cut, so we had to filter defects by where on the hide they sat.

The second wall was that there was no foundation at all. The external data scientists before me had left no pipeline, no code, no architecture, not even GPU quota on Azure. After Vodafone, where every next step is pre-decided, I had landed somewhere the steps were not even defined. Honestly, it was thrilling, like a bird out of a cage.

Building everything from zero

What happened: With no pipeline and no team, every decision was mine, from how the operator would annotate to how the model would train and serve.

The fork: Wait for structure and specialists, or build the whole thing myself.

My call: Build it. I read every Azure manual I could, and even started AppAcademy's full-stack program to design the operator's interface, though I was no front-end developer. I built the annotation flow around the operator: a projector over the hide with a touch surface so he could mark defects the way he naturally worked, plus a "useful area" he drew first so the model skipped unused regions (this solved the contextual-defect problem and saved him time). From there: annotated images to Azure, tiled, batched into CVAT for external annotators to tighten the operator's loose labels, internal QA, training on Azure, and a GPU endpoint the machine called over a simple FastAPI request, with results on the leather in under a minute.

What it led to: A working end-to-end pipeline, built alone. When the UI crashed on demo day and the robotics company's consultants could not fix their own React front-end quickly, I did, I had to: those AppAcademy days paid off. I am no front-end developer, but I have written enough C, MATLAB, Python, and Java to read any code. That is when I realized we might not need a dedicated front-end hire, and later, not even a dedicated back-end one.

Accepting the limit: we were playing the wrong game

What happened: The client was thrilled. I was not. They brought a new leather with a strange reflective finish that broke every lighting setup and showed defects in the image that did not exist on the hide. Annotating it cleanly was impossible, and false positives were everywhere. To me it was a failure.

The fork: Keep fighting for a perfect detector, or question whether perfection was even the goal.

My call: I stopped thinking about my own scorecard and asked the operator how the shipped machine was actually working for him. To my surprise, he was happy: yes, false positives, but it was already making him faster and more accurate by clearing the obvious defects off his plate. Even he had a 30% miss rate. The goal was never a perfect detector, it was a faster process. So I stopped fighting and used the model's confidence: above 90% as "confident", below 90% as "check this", so the operator's attention went straight to the uncertain calls.

What it led to: Not every confident call was right, but most were, the error rate stayed well below the operator's own, and the machine was almost twice as fast as the fully manual process. And that was a prototype, with a slow back end, a slow factory network, and a deliberately slow (for safety) robotic arm. The production machine would clear 10 seconds easily.

Impact

The prototype did its job: it convinced the client to commit to the final machine, a 9-camera system with on-board GPUs for true real-time detection. What sealed it, I think, was not only the technology. By demo day I was as strict about finding defects as the operator was, and he could feel that I understood his pain, because I had spent months doing exactly his job. He joked that if the project ever failed, I could come work as an expert annotator.

Takeaways

Know the real objective. The goal was a faster operator, not a perfect detector, so confidence-tiering beat chasing zero false positives.
A greenfield with no senior, no pipeline, and no playbook is not a blocker. It was the most fun I have had, and a generalist who can read any code (even the front-end) ships it.
When I hear someone sell AI automation for a process they have never performed themselves, I get skeptical, and a little tempted to go have a word with their clients 😊