TRAIL TUTOR

TrailTutor AI is an outdoor learning companion that encourages users to take a break from screens and engage with the real world. The AI generates one short outdoor mission based on the user's age, location, topic of interest, and available time. Users then complete the activity outdoors, returning only to reflect on their experience. This approach aims to keep the screen interaction to a minimum, with the real world being the primary focus of the learning experience.

The AI component of TrailTutor is built using an open-weight model called Gemma, which allows for greater flexibility and the ability to fine-tune the model for specific tasks. The developers tested the model using the Tinker tool to apply LoRA fine-tuning to Qwen/Qwen3.5-4B. This process resulted in a 5 percentage-point improvement in the structured output score, as well as a 13.4% reduction in average latency.

To deploy the application, TrailTutor was built using FastAPI and hosted on Render. The computationally heavy Gemma inference is handled through NVIDIA's hosted API, while the frontend and public API endpoints are also hosted on Render. The lightweight design of the web service enables quick access for judges and users to try the application.

By using open models, TrailTutor benefits from the ability to inspect, fine-tune, and experiment with the AI components. This approach allows for more meaningful experimentation and a deeper understanding of the model's behavior and limitations. The authors also acknowledge the importance of evaluating both the model and the evaluator itself, as a bug in the evaluation process initially led to an inflated improvement score.