PocketTrail: open-source AI that makes the screen the shortest part of a walk

PocketTrail is a new open-source AI project that transforms a brief activity request into a brief plan for a nearby green space. Developed during the first week of the Hacktoberfest Open-Source AI Challenge, this prototype uses Codex to select an activity, locate an appropriate park, and schedule a return trip all within a single screen. The initial release focuses on three parks in Paris. Users can replace these with parks they know and provide source links.

When PocketTrail receives an activity and budget, it employs a multinomial Naive Bayes model running in the browser to parse English activity intents. The model recognizes four intents: slowing down, moving, observing nature, and spending time together. The model was trained on 16 developer-authored phrases and runs entirely in the browser, demanding no third-party dependencies or remote inference.

The planner component ranks the grounded park pack and allocates time for the return journey. This separation allows for transparent inspection of the model's interpretation of requests and the arithmetic that prevents suggesting a plan that exceeds the time budget before the return. The code is open-source, and users can modify examples, inspect weights, replace venues, or override inference without sharing their data with a model provider.

Despite its simplicity, PocketTrail demonstrates the potential of open innovation in AI. Every model parameter can be regenerated from the included examples, and there's no paid model service, remote inference, or data transfer. However, the project has limitations, including straight-line distance estimates, unverified opening hours, weather, path accessibility, and entrance directions.

Future work could involve testing real user consent, improving multilingual evaluation, and adding better grounded venue access and route estimates.