WildHarvest AI: 100% Offline Campfire Recipe Generator using Gemma 7B

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

What I Built

I built WildHarvest AI, an offline-first campfire recipe generator designed for hikers, campers, and foragers. You simply input the ingredients you just foraged in the wild (e.g., wild mushrooms, bamboo shoots) and the app generates 3 safe, easy-to-cook recipe options using only your found ingredients and basic camping supplies (like a mess kit, salt, and water).

It perfectly fits the "Touch Grass" theme because it encourages people to get out into the woods, forage responsibly, and cook outdoors, all while bringing a smart chef along in their backpack.

Demo

Code

dimaswahyu-official / WildHarvest-AI

Collecting what you see in wild and i'll make a recipe's for that using Gemma:7b

🌲 WildHarvest AI

WildHarvest AI is a 100% offline, local-first campfire recipe generator built for hikers, campers, and foragers. This project is an official submission for the Hacktoberfest 2026 DEV Challenge (Week 1: Touch Grass).

When you are deep in the woods, you don't have internet access. This application solves that by running a local AI model to turn your foraged ingredients into safe, actionable campfire recipes without needing a single bar of cell service.

✨ Features

• 100% Offline Inference: Powered by Google's open-weight Gemma model running entirely on your local machine via Ollama.
• Strict Ingredient Bounding: The AI is highly constrained to only use the ingredients you actually found, plus basic camping supplies (water, salt, oil, etc.). No hallucinated ingredients!
• Campfire Ready: Recipes are designed to be cooked outdoors using basic mess kits or portable stoves.

🛠️ Tech Stack

• Frontend: Next.js (App Router), React, Tailwind CSS
• AI Integration:…

View on GitHub

(Note: I pivoted the project from a dessert app to a foraging app during the hackathon to better fit the theme, but the core engine remains the same!)

How I Built It

I built the frontend using Next.js (App Router) and Tailwind CSS. The core intelligence is powered by Google's open-weight Gemma (7B) model, running entirely locally on my machine via Ollama.

To make the AI act like a strict campfire chef, I created a custom Next.js Route Handler that talks to the local Ollama API. I set the model's temperature to 0.1 and wrote a highly constrained system prompt to prevent the AI from hallucinating ingredients the user didn't actually find.

Why Does Open Innovation Matter?

For a foraging and camping app, open innovation isn't just a nice-to-have; it is the only way the app functions.

When you are deep in the woods or at a campsite ("touching grass"), you rarely have a stable internet connection. If I had built this using a closed, cloud-based API like OpenAI or Claude, the app would be completely useless on the trail. By using an open-weight model like Gemma running locally on a laptop via Ollama, the entire AI inference happens offline. Open innovation made it possible to bring powerful AI into the wilderness without needing a single bar of cell service.

My Agent Session

(N/A - I built the logic directly based on local LLM experimentation)

Prize Categories

I am submitting this project for the following category:

• Best Use of Gemma: The project relies entirely on Google's open-weight Gemma model running locally for its offline recipe generation logic.

Comments

[ai_adam]

Thấy bài này rất thú vị — chạy Gemma 7B hoàn toàn offline cho use case outdoor là hướng tiếp cận đúng đắn. Mấy bác có chia sẻ thêm về pipeline quantization không? Dùng GGUF q4_k_m hay q5_k_s? Trên device gì (Pi 5, Jetson, laptop cũ) thì latency accept được ~2-3s/token?

Mình từng test Gemma 7B quant 4-bit trên Pi 5 8GB — prompt engineering cho recipe generation tốn khá nhiều context window vì phải nhét system prompt + few-shot examples + ingredient list. Có thử sliding window attention hoặc KV cache quantization không? Hay đơn giản là cắt bớt few-shot, dùng structured output (JSON schema) để giảm token output?

Còn phần "touch grass" — offline-first AI thực sự enable được use case này. Cloud LLM thì không có signal ở rừng. Mình thấy architecture pattern này (local inference + RAG nhẹ cho knowledge base) sẽ bùng nổ cho edge AI năm tới — found it via LabAgent, site: labagent .tech