WildStep AI - Local AI. Real-world missions. Zero scrolling.

NAINISH JAISWAL — Thu, 08 Oct 2026 20:49:07 +0000

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

WildStep AI — Local AI. Real-world missions. Zero scrolling.

WildStep AI turns AI into a reason to leave the screen, not stay on it.

Instead of another chatbot session, WildStep gives users short outdoor missions such as:

“Find something bright red in nature.”

Users go outside, enter Voice-First Field Mode, follow the mission, capture photo evidence, and record a live GPS trail. WildStep combines local-first AI verification, voice guidance, haptics, GPS, PWA/offline support, and a lightweight outdoor journal into one experience.

Key features:
• Local photo-evidence verification pipeline
• Voice-first outdoor Field Mode
• Client-side live GPS trail
• PWA / offline shell
• Haptic feedback
• OLED Pocket Mode
• One-mission-at-a-time outdoor HUD
• Local-first privacy architecture
• Deterministic demo runner
• 133 automated tests

The core idea:

Prepare on screen. Put the phone away. Touch grass.

Demo

WildStep AI demo:

drive.google.com

The demo shows the complete outdoor interaction:

Mission → Field Mode → Voice Guidance → Photo Evidence → Local Verification Pipeline → GPS Trail → Pocket Mode → Completion.

The photo-evidence interaction uses a real reusable Wikimedia Commons image of a Common Poppy, with attribution preserved.

The recorded environment did not have Ollama installed, so the local AI verification failure is shown honestly rather than fabricating a Gemma result.

GitHub:
https://github.com/Nainish3115/WildStep-AI

WildStep AI is an open-source MIT-licensed project containing the complete implementation, including the Python backend, Field Mode, photo verification pipeline, GPS trail engine, voice interaction, PWA shell, security hardening, deterministic demo runner, automated tests, and attribution documentation.

WildStep is a lightweight, local-first outdoor AI application built around a Python backend and browser-native capabilities.

The architecture connects the browser/PWA to the WildStep server and a local AI inference interface. Camera input, speech synthesis, geolocation, local storage, vibration, and PWA functionality are handled through browser-native APIs.

The GPS trail is entirely client-side. Coordinates are not sent to external map providers or cloud services.

For photo verification, the browser sends the selected evidence photo and mission information to the /api/check endpoint. When the local model runtime is unavailable, WildStep fails safely instead of inventing a result.

I also built a deterministic demo runner using the real production components with controlled fixtures for reproducible demonstrations.

The project was continuously validated through automated testing, manual API testing, security checks, data-flow audits, and deterministic demo execution.

The final implementation contains 133 automated tests.

AI should not only help us consume more information. It can also help us experience more of the physical world.

WildStep explores what happens when open and local AI is used to encourage people to leave their screens and interact with their surroundings.

Open models make it possible to experiment with local inference, multimodal AI, privacy-preserving workflows, offline experiences, and custom AI behavior without building the experience around a cloud-only API.

WildStep changes the question from:

“What can AI make me look at?”

to:

“What can AI encourage me to go experience?”

The goal is simple:

Use AI to get people outside.

AI coding agents were used throughout development as engineering collaborators for repository analysis, implementation, debugging, testing, security hardening, UX refinement, documentation, and demo preparation.

The generated work was continuously reviewed and validated through automated tests, manual verification, deterministic execution, and repository audits.

A key example was the photo-verification demo. When the local Ollama runtime was unavailable, WildStep returned its genuine HTTP 503 response instead of fabricating a Gemma result.

That became an important engineering principle for the project:

AI assistance should accelerate development — not replace verification.

Best Use of Gemma — WildStep is designed around local multimodal open-weight AI through Ollama and the gemma4:e2b model for photo-evidence verification.
Best Use of Temporal — WildStep includes a durable workflow integration using Temporal for persistent execution and recovery.

DEV Community: NAINISH JAISWAL

WildStep AI - Local AI. Real-world missions. Zero scrolling.

Demo