HeySalad ®, Smart Kiosk

Inspiration

Making the world healthier, one meal at a time. We wanted to create technology that inspires people to eat nutritious food while helping restaurants reduce costs through smarter operations. Food should be both delicious and good for you - AI can help people discover healthy options they'll actually love while optimizing business efficiency.

What it does

HeySalad Smart Kiosk runs GPT-OSS 20B locally on a Raspberry Pi 5 to provide personalized menu recommendations. Customers connect to our WiFi, chat with "Liv" (our AI assistant) about their preferences, discover healthy options, place orders, and receive internet access. The system promotes better nutrition choices while creating a sustainable business model for restaurants through WiFi monetization and reduced operational costs.

How we built it

Hardware: Raspberry Pi 5 configured as a WiFi access point with captive portal AI: GPT-OSS 20B via Hugging Face Hub with local processing Frontend: Responsive JavaScript interface with floating chat design Backend: Python Flask server with intelligent fallback responses Database: Firebase for analytics, local SQLite for offline operation Network: Custom captive portal using iptables and hostapd

Everything runs offline-first - customers get AI recommendations without internet dependency.

Challenges we ran into

Getting GPT-OSS running smoothly on Raspberry Pi hardware required optimization work. iOS Safari doesn't support voice recognition, so we built dual interfaces. Captive portal detection varies across devices, requiring multiple endpoint handlers. Python dependency management on Raspberry Pi OS needed workarounds with virtual environments.

Accomplishments that we're proud of

Successfully deployed GPT-OSS 20B on affordable edge hardware under £200. Created a unified interface that adapts to device capabilities. Built a complete business solution addressing real needs. Achieved 15W power consumption versus 150W+ cloud alternatives. Developed production-ready captive portal handling device management and analytics.

What we learned

GPT-OSS models work effectively on Raspberry Pi with proper setup. Progressive enhancement is key - build core functionality first, add advanced features for capable devices. Captive portal mechanics are complex but manageable. AI prompt engineering requires careful crafting for helpful responses. Technology can create win-win scenarios - better customer experience plus business value.

What's next for HeySalad Smart Kiosk

Q1 2025: Fine-tune GPT-OSS on nutrition data and add multi-language support Q2 2025: Integrate ESP32 sensors and BLE beacons for inventory tracking
Q3 2025: Deploy multi-location network with federated learning Q4 2025: Expand into autonomous food service with computer vision and robotics

Building a network of intelligent kiosks that promote healthy eating while reducing costs across train stations, campuses, and communities worldwide.

Built With

• firebase
• hugging-face
• open-ai
• python
• raspberry-pi