I built a hiking guide that runs on a laptop with no internet — meet trailbrief

utkarsh sonawane — Sat, 10 Oct 2026 20:25:55 +0000

What I built

trailbrief is a small Python CLI that writes you a personalized hiking briefing. You tell it the trail — name, distance, elevation gain, location, season, your experience level — and it hands back a practical plan: an overview, a pacing plan, a season-aware pack list, watch-outs, and a Leave No Trace reminder.

The twist: the briefing is written by Gemma, running locally through Ollama. No accounts, no API keys, no data leaving your machine. And because the whole point is getting outside, there's a proper offline fallback — when no model is reachable (read: you're already on the trail with zero bars), a built-in template engine still produces a genuinely useful briefing instead of an error message.

Repo: https://github.com/sonawaneutkarsh/trailbrief

Demo

Here's a real run — offline mode, since I don't have a GPU farm in my dorm:

$ python3 -m trailbrief \
    --name "Ricketts Glen Falls Trail" \
    --distance 7.2 --elevation 1200 \
    --location "Benton, PA" \
    --season fall \
    --experience intermediate --offline

# Trail Briefing: Ricketts Glen Falls Trail
*Benton, PA · 7.2 mi · 1,200 ft gain · Fall · Moderate*

## Overview
A moderate day out: 7.2 miles and 1,200 ft of gain asks for steady pacing and a real lunch stop. Expect about 4.2 hours of moving time, plus stops.

## Pacing plan
Start no later than 8 AM — with 4.2 hours of moving time plus breaks, you want a buffer before dark.

## Pack list
- Insulating mid-layer — temps swing 20°F+ through the day
- Headlamp with fresh batteries (shorter days sneak up on you)
- Bright colors — it's hunting season in many areas

## Watch-outs
Daylight is the constraint: sunset comes early and temperatures drop fast once the sun dips. Leaf-covered rocks are slippery.

With Ollama running (ollama pull gemma3, then drop the --offline flag), the same inputs go to the model instead — with a twist I'll get to below.

How it works

The architecture is deliberately boring, in the good way:

Deterministic core. Difficulty and moving time are computed from your numbers with a Naismith-style rule (2 mph + 30 min per 1,000 ft of gain). No vibes, just arithmetic.
Grounded prompt. Those computed numbers are baked into the prompt, and the model is explicitly told to use them and not invent trail features, facilities, or water sources. This is the part most "AI hiking apps" get wrong — a chatbot happily hallucinating a water fountain at mile 6 is how people get hurt.
Local inference. The prompt goes to localhost:11434 — Ollama, plain HTTP, stdlib only. Default model is gemma3.
Honest fallback. Model unreachable? The template engine takes over with season-aware packing (fall gets headlamp + hunter-orange reminders, winter gets microspikes), elevation-aware safety notes, and experience-level tips.

The full test suite (12 tests, python -m unittest discover -s tests) covers the difficulty bands, prompt grounding, fallback content, and the Ollama HTTP layer with mocks — no server needed to verify.

Why open innovation matters for this build

This project only makes sense because the AI is open:

It works where the internet doesn't. A cloud API can't brief you on a ridgeline with no signal. A 3–9B open-weight model on your laptop can. The "touch grass" theme isn't decoration here — offline capability is the entire point, and only open weights make it possible.
Your location stays yours. Where you hike is personal data. With local inference, your trail plans never touch a server you don't control.
It costs nothing to run. No per-token billing for a hobby tool you use twice a month. The economics of a closed API would kill this project; open weights make it free forever.
Models are swappable, not load-bearing. Don't like Gemma's briefing style? --model llama3.2 and you're done — no code changes, no renegotiating with a vendor. The deterministic core (difficulty math, grounding, fallback) stays identical no matter which weights you plug in.

A closed model could write a prettier paragraph. It couldn't do any of the four things above. That's the whole argument.

What I'd do next

Trailhead weather overlays, GPX import so distance/elevation come from the file instead of your memory, and a --compare mode that briefs the same trail with two models side by side. All doable without touching the core.

Prize categories

Overall
Best Use of Gemma — Gemma (via Ollama) is the default briefing engine; the prompt, grounding strategy, and model-swap design are built around it.

Built October 10, 2026 for the Hacktoberfest Week 1 DEV Challenge ("Touch Grass"). Tags: #devchallenge #hf26challenge