FoodieCompass : The No-Screen Walk-to-Eat Guide

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

What I Built

FoodieCompass

It's a simple app that allows user to tap just one button to find out their breakfast/lunch/brunch/dinner place , then just pockets the phone and walk around. *Work best for students and 9-to-5 workers who does not own their transport *

Why I Built

When people want to eat out, most of us just get trapped in endless "review paralysis" no matter using Yelp , Google Map or someone blog , sometimes maybe 15 minutes walking just turn into an empty and disappointing restaurants.

How it solved

1. User only need to open the app -> choose their preferences (NO TYPING AT ALL) , the UI will shows a single analog-style compass arrow pointing directly toward the single highest-sentiment spot.
2. To reduce the frequency of taking out phone just to make sure users' current positions, a haptic vibrations and short audio cues will support them at intersections
3. An open-weight Small Language Model (SLM) running directly on-device analyzes raw community reviews, detecting signals of genuine crowd buzz, outdoor seating, and authentic quality while filtering out tourist traps and empty venues.

Core Features

1. "One-Tap Stroll" Selector
Simple presets: "Quick Bite (5 min)", "Dinner Stroll (12 min)", "Outdoor/Street Food Walk (15 min)".
No typing, no infinite scrolling, no search filters.

2. Open-Weight Auto-Filtering Engine
Scans candidate eateries within walking distance.
Open SLM classifies places based on:

Crowd Authenticity: High local turnover vs. dead restaurant risk.
Signature Dish Consensus: What item diners actually rave about.
Outdoor Vibe: Patio seating, street tables, or walkability.

3. The Compass UI
A clean, full-screen directional needle that rotates in real time using the device's magnetic compass sensor even user is offline.
Shows only:
Destination name & distance in meters (e.g., 320m • 4 mins walk).
Target dish pill (e.g., "Crispy Roast Pork").
1-sentence AI verdict: "Consistently lively patio; locals praise the handmade dumplings."

4. "Eyes-in-Pocket" Sensory Guidance
Haptic Pulse: Phone gently vibrates when the user is walking in the exact heading of the food spot.
Proximity Buzz: Distinct vibration sequence when within 40 meters of the destination.
Audio Chime : A subtle chime or whisper cue: "You have arrived. Look to your right."

Demo

foodiecompass.vercel.app

Code

123EFD / Hacktoberfest

Food Compass - Simple Website for foodie , a few click then start your hunting

How I Built It

Open-source AI

• Open-Weight Model : SmolLM2-360M-Instruct (Hugging Face) — an ultra-compact, high-efficiency open-weight Small Language Model (SLM)
• Local Inference Framework : Transformers.js (v3) powered by ONNX Runtime Web with WebGPU hardware acceleration and WebAssembly (WASM).
• Open Knowledge & Data Engine : OpenStreetMap (OSM) / Overpass API — an open, decentralized geographic database acting as the community knowledge harness for local eateries, replacing closed commercial map APIs.
• Open Client Edge Infrastructure : Packaged via Capacitor and Vite PWA (Workbox Service Worker), allowing the entire AI inference runtime, compass math, and synthesizer code to run completely offline on client hardware.

How is my project built around it?
Compared to traditional food apps which require you to read many reviews and choose what's the best for you , FoodieCompass reduce this feature unnecessary via open-source AI.

1. The open models digests raw venue data (seating, cuisine and dietary tags etc.) and filter out hundreds of potentials choices into one single high-confidence destination.
2. Once the destination is selected, local hardware sensors (your magnetometer chip inside your phone) will take charge of the GPS coordinates as well as the notification, vibration and audio assistance.

Why Does Open Innovation Matter?

• This app is built for those who prefer location privacy. When using commercial platforms ( Google Maps , Yelp), every GPS coordinate , walking path and dining craving is transmitted to commercial ad servers to build behavorial ad profiles. This open-weight AI running directly on the device without search query and location packet leaves user's phone.

• Zero-cost : Running an open-weight model locally costs $0.00 in cloud fees, making this tool completely sustainable and free for anyone in the world to self-host or install.

My Agent Session

The full interactive agent transcript of this build session is documented here:

• View Full Agent Session Transcript (session.jsonl)

Submitted by @kas_6156c2c3d9f90