Developers are increasingly looking for ways to reduce screen time, and one Hacktoberfest submission is tackling this by using browser-based AI to encourage physical activity. Walking Challenge AI is a lightweight web application that generates dynamic, real-world tasks for users to complete during a walk, leveraging open-weight models running entirely on the client side. Built for the Hacktoberfest Open-Source AI Challenge Week 1, the project demonstrates how local inference can power simple, privacy-focused utilities without relying on cloud APIs or backend servers.

Local Inference With Transformers.js

The core technical achievement here is the deployment of a Qwen2.5 Instruct model directly in the browser using Transformers.js. The app supports both WebGPU acceleration and WASM fallbacks, ensuring compatibility across different environments, though the initial model download requires approximately 1.2 GB. Users configure their session by selecting a walk duration between 10 and 60 minutes, a location type such as a park or city, and a difficulty level. The model then processes these parameters to generate a list of specific challenges, such as identifying tree species or observing the sky for a set duration.

Handling Model Unpredictability

The developer, tamim_ts1, encountered significant hurdles with smaller parameter models during the build process. An initial attempt using a 0.5B parameter model failed to follow instructions, often echoing the prompt rather than generating new content. To mitigate this, the application includes a robust validation layer and a hand-written bank of fallback challenges. If the local model returns unusable output, the app seamlessly switches to pre-defined tasks, ensuring the user experience remains consistent even if the AI generation fails. This architecture highlights a common pitfall in edge AI: small models are often unreliable for structured output without careful engineering.

Privacy and Cost Efficiency

Because the inference happens locally, user data regarding time, location, and activity never leaves the device, addressing privacy concerns for an app that tracks where people physically go. There are no API keys, no backend costs, and no per-request fees, making it an ideal candidate for student projects or low-budget deployments. The codebase is designed for extensibility, with the model configuration isolated in a single line of code within src/ai.js, allowing developers to easily swap in fine-tuned models tailored to specific campuses or cities.

Key Takeaways

  • Walking Challenge AI uses Qwen2.5 Instruct via Transformers.js for browser-based inference.
  • The app requires a ~1.2 GB initial download but caches the model for subsequent uses.
  • A fallback mechanism with hand-written challenges ensures reliability when model output is poor.
  • The project is open-source and built specifically for the Hacktoberfest Open-Source AI Challenge.

The Bottom Line

This project is a refreshing reminder that AI doesn't always need to be complex or cloud-dependent to be useful. By solving a simple human problem with local inference and a solid fallback strategy, it offers a practical template for building privacy-first, zero-cost AI tools.