Winning Demo Day was an great conclusion to our project, especially because the votes came from
our peers. Shoutout to my team: Jens, Kas and Shashank.
Looking back It’s funny that a simple assigned chat by our professor with peers would lead to a
synergistic collaboration. I connected with Shashank, we got along quite well and I noticed his
technical background and we decided to pair up. Jens and Kas joined us organically during the lecture
break, completing our team. During brainstorming for the “Short Project Proposal” we got inspired by
the course focus on digital disruption and chose to tackle operational inefficiencies in auto insurance.
Our progress wasn’t linear: after submitting our short proposal, we learned that feedback was only
given if a group’s idea needed steering or correction. Lacking external pressure because we had none
we assumed our concept was solid and just leaned back. A bit too much actually and our progress
stalled. When we realized we were losing momentum, we decided to gather up in person to make a
plan.
Shashank came prepared with a strategic plan, dividing the project into distinct phases and assigning
each with specific slide responsibilities. We also made a strict rule. If you get stuck on a task, tell the
group 24 hours before the internal deadline. Besides that we started hanging out between classes,
which contributed to the team spirit from my view. This made me realize a lesson about teamwork:
structured planning paired with informal moments allowed our team to quickly ask for help and solve
problems collectively rather than getting stuck in isolation.
Shashank took the lead on the MVP. I took responsibility for slides 11, 12, 13, and 15 (My objective
was to design a testing protocol that proved the concept while ensuring the AI remained safe, legally
compliant). Our model was great at predicting things, like spotting a damaged bumper from a few
photos. But letting it handle human judgment? We could not allow an algorithm to assess injuries.
We built hard guardrails. Any text mentioning an injury instantly triggered a human-in-the-loop
fallback.
Then I tried to build a test dataset for our pitch (Absolute nightmare). There is no public dataset out
there with the exact wide shots, close-ups, and license plates we needed. Let alone datasets that also
have verified repair invoices attached. Which eliminated our ability to test cost accuracy. I mixed real
damage photos with random internet pictures of undamaged cars in the exact same color. Our AI
instantly flagged the whole batch as fraud and rejected them. Great for proving our fraud-detection
works. Terrible for my testing plan.
Applying a flat price per car part is economically flawed anyway. A BMW bumper costs way more than
a Toyota one and labor rates are all over the place depending on the garage. I texted the group on
WhatsApp and we just pivoted. We dropped the financial accuracy test. We didn’t have the
proprietary data anyway. We narrowed the test down to part accuracy, damage severity, and safe
escalation and tested on a new data set (not on the previous mentioned flawed data set). We put in
our limitations that insurers have to load their own price tables. The true competitive advantage for
an insurer is not the AI model itself. It is the proprietary feedback loop generated by integrating their
specific financial data.
With the testing parameters and market strategy set, I audited our legal compliance. Our initial
design failed to explicitly notify users that an algorithm was processing their data. Operating under
strict transparency and privacy standards, that is a massive liability. I noted that we need to inform
users explicitly about an AI model processing their claim. We had to deliver a compliant tool, not just
a capable one.
Then Shashank introduced an angle I hadn’t anticipated: adversarial testing. He actively tried to break
the system. He deployed prompt injections. He embedded malicious instructions inside uploaded
photos, literally commanding the AI to approve €5,000 payouts. The system blocked 19 out of 19
attempts. Developing a viable AI product is not just about ensuring it functions under ideal
conditions. It must fail safely under malicious conditions.
This entire project flipped how I view digital strategy. I previously assumed deploying AI meant
possessing the smartest algorithm. I now recognize that AI simply lowers the cost of expertise.
Anyone can access an API to analyze a dented car. The real value is how a business orchestrates
human judgment, builds trust, and locks that technology into an ecosystem nobody else can copy. If a
competitor can just buy the same AI tool, you don’t have a business model. You just have a feature.
Demo Day provided one final insight. While I was proud of our project regarding an insurance
platform, I found myself excited by another group’s project calculating/ managing to view
macronutrients from restaurant meals. My enthusiasm originated from my own personal interests
with eating healthy (which for me requires insight into my macro nutrients intake). To me that was a
strong reminder, because I was instantly excited me about their project. So choosing a problem you
actually care about remains the true driver of innovation.
