SixtSeven

Our landingpage
Our carboard vendingmaching with built in barcode reader and key dropper
The landingpage where the users book create bookings to get the QR Code
The QR Code to use for the booking
The Landingpage on the vendingmachine
Our favorite bird mascot doing its thing :)
Personalized car suggestion
Individual insurance recommendation
Key release page
Summary page

Inspiration

In our fast paced society, flexibility and efficiency have become invaluable. Every process is being optimized and the rental car industry is no different. Long queues at the car rental counter can make the excitement of the trip fade into paperwork and waiting. When you’re in a hurry, the last thing you want is to engage in a full-blown conversation with an AI voice agent. That’s why we set out to extract valuable information without requiring any additional user interaction resulting in less friction for the customer. We want to streamline the physical hand-off to be as fast as possible, while elevating the customer experience through extreme customization.

What it does

Our goal was to emulate a human service agent as closely as possible while staying out of the users way and keeping the overall experience smooth and effortless.

The user arrives with a confirmed booking and a QR code containing their booking ID on their phone.
They approach our vehicle-key vending machine and start the process by scanning their QR code with the built-in scanner.
In parallel, a local on-device machine-learning model analyzes the user in real time - detecting the number of travelers, whether they have luggage, and other useful contextual attributes.
Using these insights, the system proactively suggests relevant vehicle upgrades, insurance options, and helpful add-ons - all without forcing the user to answer unnecessary or annoying questions.
Throughout the experience, our friendly, talkative bird mascot guides the user with clear, cheerful instructions.
After the user accepts or declines the personalized upgrade suggestions, the machine dispenses their key through the integrated key retrieval system.

How we built it

The brain - In the age of AI, most applications require a complex logic with storage and compute issues. We opted for a lightweight and fast computable decision tree to customize the recommendation system for all customers.
The eyes - If you’re rushing, having a voice agent insist on a conversation sub-optimal. Our workaround is to avoid the talking altogether and let the system gather key info on its own, with no effort from the user. To achieve this, we needed a tiny, real-time computer vision setup that could run on a consumer device’s CPU (kiosk) and still accurately extract features like person count, luggage count, or even whether someone is wearing a Hawaiian shirt from a video stream. With data privacy top of mind, we designed the entire pipeline to run locally on the device, ensuring that no image data is ever sent to an external AI provider. At first glance, a “Hawaiian-shirt classifier” might sound like a joke - but we picked it deliberately. It’s a simple, vivid example that showcases how easily you can train models from scratch on a wide range of features.
The voice - We used ElevenLabs for natural voice synthesization.
The face - We built the Interface of our Key vending machine using SvelteKit and tailwindcss. Our friendly companion mascot bird makes the process very pleasant and likable
The hands - This time around, we wanted to build something physical as well. So we used some scrap cardboard, an esp32, a servo and too much hot glue to build a simple vehicle key dropper which we integrated to our system over the network.

Challenges we ran into

Finding and training a computer vision segmentation approach that provides valuable insights while still working with our given resource and privacy targets
Presenting offers to users in a digestible way without being too intrusive
Our API Keys ran out of quota

Accomplishments that we're proud of

We managed to build a fully working system including a machine learning pipeline, several API integrations, web deployment, and hardware components
Even though we struggled quite a bit with outlining and specifying our goals, we managed to develop a cohesive and consistent concept which worked out great in practice
While it's tempting (and easy) to stuff AI into anything nowadays, we managed to keep it sensible and use it to an extent that actually provides value while still preserving privacy and being mindful of resource usage.