This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass
The Backstory: Why I Needed to "Touch Grass" More Than Anyone
To be completely honest with you all, when I saw the theme for Week 1 was "Touch Grass", I actually laughed out loud.
For the past several months, my entire life has basically been lived inside a terminal and an IDE. Between spending hundreds of hours building web components for Google Summer of Code (running tests, debugging PRs, staring at screen pixels until 2 AM) and college classes, my screen time was through the roof. My posture was terrible, my eyes were constantly strained, and "going outside" meant walking 5 steps from my desk to grab a glass of water.
Every programmer knows that feeling: you're stuck on a bug or drained after a long coding marathon, and your brain just feels fried.
The best cure for this isn’t scrolling Twitter or watching YouTube. It’s literally stepping outside, walking under trees, feeling the breeze, and moving your body. Science backs this up too—just 20 minutes of outdoor walking resets your mental fatigue, boosts problem-solving, and clears your head.
But here’s the problem:
Whenever I tried going for an afternoon walk around my campus or neighborhood, it was either blistering hot on the open roads or I'd end up pulling out my phone to check weather apps, maps, or notifications.
Most weather apps just tell you: "33°C, Sunny." That's completely useless on the ground! Anyone who walks knows that walking 100 meters on black asphalt feels like being baked in an oven (easily 40°C+ radiant heat), while taking a brick path under green trees right next to it feels like an oasis.
I didn't want another bloated app that keeps my face buried in a screen. I wanted an app that does its job in under 10 seconds so I could put my phone away in my pocket and actually walk.
That’s why I built MicroClime.
What I Built
MicroClime is an open-source, 100% offline microclimate companion designed to help you discover the coolest, most comfortable walking routes in your neighborhood—with practically zero screen time.
The design philosophy is dead simple: the app should never distract you from being outside.
Here’s how you use it when you step out the door:
-
Quick Tap (takes ~5 seconds): While you're walking, you tap 4 quick buttons:
- Surface under your feet: Grass, Brick, Concrete, Asphalt
- Shade level: 1 (Baking sun) to 5 (Full tree canopy)
- Breeze: Yes / No
- How hot it actually feels to you: 1 (Chilly/Pleasant) to 5 (Scorching)
- Instant Local Prediction: After logging just 5 to 8 quick spots around your neighborhood, Prior Labs TabPFN-3 evaluates all 40 possible combinations of surface, shade, and breeze on the fly.
- Natural Human Advice: Instead of forcing you to squint at complicated charts or data tables in the bright sunlight, Google Gemma 2 (2B) translates the top predictions into a 2-sentence outdoor walking tip (e.g., "Stick to the shaded grass perimeter behind the academic blocks; avoid the open asphalt main road right now").
- The Screen-Off Prompt: The app explicitly tells you: > "Goal Accomplished: Screen time over. Put your phone in your pocket and enjoy the walk!"
Demo
(The interface has big, thumb-friendly buttons designed specifically so you can log a spot with one hand without even stopping your stride. It runs locally and opens directly on mobile browsers over local Wi-Fi!)
Code
I've made the entire project open-source:
🌿 MicroClime: Offline Urban Heat Scout
Touch Grass, Not Screens.
Built for the Hacktoberfest Open-Source AI Challenge (Week 1 / 2026).
Targeting: Best Use of TabPFN & Best Use of Gemma.
📖 Overview
MicroClime is an open-source, offline-first outdoor companion that helps people escape indoor screens and safely navigate outdoor urban heat islands.
During outdoor strolls, runs, or campus walks, users log a few sensory data points with a single tap (surface ground type, canopy shade, breeze, and perceived heat).
In under 5 seconds:
- Prior Labs TabPFN-3 (tabular foundation model) evaluates all 40 microclimate permutations to forecast the coolest walking corridors and identify dangerous heat traps.
- Google Gemma 2 (2B) (open-weight model running locally via Ollama) translates these tabular predictions into actionable, plain-English walking navigation advice.
- The interface immediately tells the user to lock their screen and enjoy the outdoors.
âš¡ The "Touch Grass" Philosophy
The challenge…
GitHub Repository: https://github.com/SujalTripathi/MicroClime
How I Built It
Building an outdoor tool meant I had to make real engineering decisions that fit the physical world. I couldn't rely on huge server clusters or heavy cloud APIs. Here’s what makes MicroClime tick under the hood:
1. TabPFN-3 by Prior Labs (The Tabular Brain)
This is where traditional machine learning usually fails. If you walk outside for 15 minutes, you only collect maybe 8 to 12 data points.
- If you try training a normal XGBoost, Random Forest, or neural net on 10 rows, it either completely overfits or fails miserably.
- You don't have time to do hyperparameter tuning or cross-validation while standing on a sidewalk!
- TabPFN-3 is a foundation model trained specifically for in-context inference on tabular data. It treats your small walk log as a prompt context and generates probabilistic predictions across every possible terrain combination in milliseconds—with zero training runs.
2. Google Gemma 2 (2B) via Ollama (The Screen-Free Translator)
Nobody wants to read raw numbers, confidence intervals, or tables while walking outside under a glaring sun.
I run Google's lightweight open-weight Gemma 2 (2B) locally using Ollama. Gemma takes TabPFN's numerical rankings and converts them into a friendly, concise, 50-word outdoor walking recommendation. It gives you the direction you need immediately, without keeping you glued to the display.
3. Streamlit & Python UI
I wanted something responsive, tactile, and fast. Streamlit allowed me to build big, accessible touch buttons that work smoothly on mobile browsers so you can log a spot in two taps.
Why Does Open Innovation Matter?
This theme really made me appreciate the open-source and open-weight AI ecosystem. If I had built this using closed proprietary APIs, the project would completely miss the point:
- Zero Signal, True Offline Autonomy: If you walk into a state park, dense tree cover, or rural trails, mobile data drops to zero. Cloud APIs (like closed LLMs) immediately break and leave you hanging. Because TabPFN-3 and Gemma 2 run on local weights directly on the device, MicroClime works seamlessly in airplane mode in the middle of nowhere.
- Real Location & Routine Privacy: Where, when, and how you walk every day is deeply sensitive personal information. Nobody should have to upload their daily physical routes to commercial cloud servers just to check microclimate heat pockets. Open weights keep 100% of your personal walk logs on your own hardware.
- No Paywalls or API Bills: Health and outdoor tools should be accessible to any student or walker. Running open-weight models means no monthly subscriptions, no token metering, and no surprise rate limits.
-
Predictability & Control: Pinning our TabPFN configuration to
ModelVersion.V3in open code gives full reproducibility without fear of an upstream vendor deprecating an endpoint overnight.
My Field Test: Actually Putting It to Work
I decided to test MicroClime the best way possible: I shut my laptop lid, put on my walking shoes, and headed outside around 1:30 PM when the sun was strongest.
Here’s what happened on the walk:
-
The Walk: I took a 15-minute loop around our area. Over the walk, I logged 6 quick spots:
- Open asphalt street (felt heat: 5/5, blazing)
- Concrete walkway beside the building (felt heat: 4/5)
- Brick pathway under light shade (felt heat: 3/5)
- Tree-lined lawn with grass (felt heat: 1/5, super breezy)
- What TabPFN predicted: Without any manual training, TabPFN-3 picked up the microclimate dynamics instantly. It predicted that grass with canopy shade would maintain an index of ~1.1/5 even as midday peaked, while asphalt remained a 4.9/5 heat sink.
- What Gemma advised: Gemma spat out a clean 2-sentence tip: "Keep to the grassy lawn corridor on the north side—the canopy shade and breeze drop the heat significantly. Avoid the central asphalt road until evening."
- The Screen Time Result: My total time looking at my phone screen over the entire 15-minute walk was less than 40 seconds combined. The rest of the time, I just walked, enjoyed the fresh air, listened to the birds, and let my brain recharge.
When I got back to my desk, my headache from hours of staring at code was completely gone.
Prize Categories
- Best Use of TabPFN: MicroClime leverages Prior Labs' TabPFN-3 tabular foundation model to perform zero-shot in-context predictions on small, real-world walk logs (<15 rows), solving the microclimate heat prediction problem without needing training loops.
- Best Use of Gemma: Uses Google's open-weight Gemma 2 (2B) running locally via Ollama to translate complex tabular predictions into natural, screen-free walking advice that actively encourages users to pocket their phone and enjoy the outdoors.
If you’ve been spending all day inside debugging or pushing commits this Hacktoberfest, take this as your sign: close your IDE, grab some water, and go touch some grass!
Top comments (0)