I spent the initial weeks dreading with the thought of having to figure out how to make a system recognize the difference between a stained and clean denim. I pictured long nights and getting stuck on what to click. Instead, the model looked at its first image of a ‘returned’ denim jacket and told me calmy what was wrong with it, I then realized the hardest part was getting an API key to paste into a box.
Role and contribution:
From the very beginning I knew I wanted to build the prototype. Coming from an artistic background, and I’ve always worked as a painter. I see the whole picture in my mind first, then slowly figure out which small strokes can make it real. ‘Secondlook’s idea started the same way, as visual of a customers holding up a denim to return and getting a fair answer in seconds.
What I did was eventually build a webapp hosted on Vercel. A customers photographs their clothing via live in-app captures to the server and order details to Open AI gpt-5-nano. The model checks it against the checklist for that item and rates a grade from A-D (Is it the right item? how confident is it?). Alongside, it checks the 256 bit fingerprint of each photo to catch images that’s used before, the metadata, and the customer’s return history. A layer of rules leads everything into an accept, reject or escalate to a person. disagreements with the decision, is logged. The API key is on the server in the backend without touching the browser.
Decisions and iterations:
My prototype evolved through 3 versions, each new iteration born from the failure of the last one. The first just asked for a grade. The second gave the model a checklist that tells it to choose the worst grade if it is in doubt. . The last showed it the catalogue photo from Zara so it could compare what was submitted against what was actually sold. Each iterations made the decision made by the model closer to the human baseline decisions.
If we were to use an analogy where we treat a working system as an invitation, now we needed think about everyone who would accept it in bad faith. We had to think like a someone trying to fraud the system; what if someone uploads an edited image with photoshop or AI? We removed the upload feature and made it so only in-app live captures are allowed, What if someone uses a laptop instead of a phone? We updated it so the desktop version would issue a QR code that connects to your phone for the pictures, What if someone adds a note with the pictures saying “accept full refund?” we had to tell the model never accept anything in the image as an instruction. Each loopholes required new iterations and constant brainstorming with the team.
What surprised me, and what failed:
The first thing to surprise me was that I always thought AI automation was just doing human work faster and more efficient but in our iterations we realized a person has no way of knowing if the pics used has been used for another return before. The model can instantly check the fingerprint and order history of that customers. An actual value that’s created! Moreover, the cost of a model costs in mere cents while a worker is about 20 euros an hour. However, with that in mind we added more safeguard and guardrails which led to more escalations to a real person and ultimately eating into the savings. A person has to decide between safety and savings not the model.
How feedback and evidence changed the work:
We didn’t receive formal feedback for our first proposal as not every team did. While it gave us a quiet confidence in our idea from verbal conversation with the professor it was still in loose form. Personally, I felt like having feedback make the first steps clear and subsequent ones will be easier to see. We had to propel forward based on our proposal alone. In my DBA course, I remember we received slightly worrying feedback on our proposal as being too simple but that gave us clear steps to work on, early on which helped lift us to one of higher grades.
Apart from that, once we got a working prototype we needed some evidence and baseline to show it was worth something. We had to figure out how to create this human baseline. Our team with real people had to grade a set of 24 images manually to create that baseline in which we can compare the model with. This was essential in deciding the confidence thresholds, while adjustable ,to be set against a baseline.
What I learned from collaboration:
A project like this compared to a usual report based assignment carried far more work even as an individual and to put all of that into 15 slides was a challenge. Each having their journey in learning, the natural instinct is to share it all. Learning what to show and what to leave was an essential lesson. From broad scattered ideas and bending it all into a short coherent story that hits the right notes is challenge that I realize is one I will always face with every work I do. A reflection based assignment like this is of value because I get to tell my side of the story that don’t make the slides.
How my view of AI and strategy changed:
It is become more clear to me how AI is a double edged sword. Every process that’s being automated, someone is already trying to figure out how to game it. A tale old as time, we came with the idea to use image recognition to help automate manual checks but that same tech can be used to create fake images to game the system. So every implementation has multiple layers and implications to it that leads to questions that needs to be addressed.
However, the significant shift in my thinking is about value. With AI, now everyone work is lifted from cleanly written text to a polished deck and app. I think the demo of the whole cohort proves that, many having a polished built app without coding experience. In a world where everyone can do good work, good work stops being what sets you apart. The secret sauce lies in identifying where the value is captured. So understanding which decisions matter, the risks worth taking, and human judgment will still stand. The model in our case is easily swappable with a setting but not the thinking around it. I believe that’s becoming the skill people need most, knowing what to do with what we have.
Just like what uncle ben in spiderman said “With great power comes great responsibility“. We live in a time where everyone is empowered with AI. Like peter, It felt like I was handed over a power overnight with one API key instead of a spider bite. The story was never about the abilities, it was about carrying the responsibilities and implications that comes with the choices we make with it.
