Kagaz: An Offline AI That Prints a Nature Scavenger Hunt So Kids Go Touch Grass

Kagaz is an offline-first tool that creates printable bilingual Hindi and English nature scavenger hunts for children. To generate a hunt, a parent inputs a city, an age range, and a month. An open-weight LLM then crafts the hunt, while a seasonal filter ensures the items are realistic for that time of year. The finished product is an A4 PDF with a checkbox grid, which can be printed and handed to the child to find items like a dry leaf with a hole, a crow on a wire, or the scent of wet earth.

The total screen time for this activity is just around 30 seconds, while the child spends 1 to 2 hours outdoors.

The app was created for the Touch Grass theme during Hacktoberfest 2026, Week 1. The choice of using open models was not based on ideology, but rather on practicality. The open route was found to be more suitable for this project, as it allows the tool to function offline, making it useful in areas with unreliable internet connections, on airplanes, or during power outages. Additionally, using open models ensures that the child's data never leaves their laptop, with no cloud involvement, accounts, logs, or training data.

The model used for Kagaz is Qwen 2.5 7B, an open-weight LLM, which was chosen over gemma2:2b for its superior multilingual output. The seasonal filter, TabPFN, and pandas were employed to select appropriate item types for the specific month. The PDF generation is handled by ReportLab, while Streamlit creates the parent-facing UI. Python 3.13 runs on a Windows 11 laptop with 15 GB of RAM.

The tool generates a JSON object with 12 bilingual items, each containing an emoji, text in English and Hindi, a hint, and a type. The safety validator ensures that the items do not contain any potentially dangerous or inappropriate suggestions, such as "touch", "eat", "pick mushroom", or "taste". After completing the scavenger hunt, the app disappears, leaving no trace of its presence.