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

Jordan Lozano — Thu, 08 Oct 2026 00:21:26 +0000

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass.

What I Built

PocketTrail turns a short activity request into a pocket-sized plan for a nearby green place. It is a new prototype built during Week 1 with Codex. Its purpose is small: choose an activity, find a place that leaves time for the return trip, save a card, then leave the screen.

The starting pack covers three Paris parks. The starting coordinates are an example, not a detected or claimed personal location. People elsewhere can replace the park pack with places they know and source links.

Demo

Try PocketTrail.

A calm activity and a 90-minute budget produced Jardin des Tuileries with approximately 36 minutes each way and 18 minutes outside. A shorter, impossible budget returned an explicit empty result. Download or print the text card and close the tab.

Code

Source, model, tests and MIT license.

Run locally with python serve.py --port 8801; run the eleven tests with node --test test.mjs. There are no third-party dependencies to install.

How I Built It

A multinomial Naive Bayes model runs in the browser and recognises four English activity intents: slowing down, moving, observing nature and spending time together. Its 16 training phrases are developer-authored and openly included. They are not survey answers or personal histories. Unknown words yield no inferred intent; an explicit activity selection overrides the model.

The model chooses an intent, not a fictional destination. A separate constraint planner ranks the grounded park pack and reserves time for the return journey. This separation makes the useful part inspectable: the model helps interpret a request, while arithmetic prevents a suggestion that exhausts the time budget before the return.

The first browser test exposed a practical Windows problem: the standard local server served .mjs files with the wrong content type. A small server now supplies the correct JavaScript type. The planner also rejects invalid coordinates, impossible budgets and empty place packs. Four independent authored holdout phrases and planner edge cases are covered by eleven passing tests. This is behavioral verification, not evidence of broad real-world accuracy.

Why Does Open Innovation Matter?

Every model parameter can be regenerated from the included examples. There is no paid model service, remote inference, account, tracking or model download. After the local files load, planning happens in the browser. A person can change the examples, inspect the weights, replace the venues or override the inference without sending their activity request or coordinates to a model provider.

A tiny open statistical model is sufficient for this narrow decision. Its limits are visible in the model card. A larger closed API would add a dependency and data transfer without fixing the missing venue information or route accuracy.

Limits and Next Steps

Distances are straight-line estimates with a walking multiplier, not street routes. Opening hours, weather, path accessibility and entrance directions are not verified. Check the venue source before going.

No outdoor trial, user study or personal experience is claimed. Future work should test consented real use, improve multilingual evaluation and add better grounded venue access and route estimates. For now, the output is a modest plan with visible assumptions.

Prize Categories

Overall challenge. No partner technology category is claimed.

AI disclosure: Codex generated the implementation, tests and this article under Jordan's direction. No human experience or feedback has been fabricated.

DEV Community: Jordan Lozano