Pranjal Manhgaye

Posted on Oct 9 Edited on Oct 9 AI-assisted

PLOT 🌿 - TabPFN on Real Frost Data Tells You to Touch Grass

#devchallenge #hf26challenge #productivity #webdev

Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass Submission 🌿

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

💡 The idea in one breath

Most weather apps keep you scrolling. PLOT does the opposite. 📵➡️🌱

You pick a place (or share your location 📍). Real weather loads 🌦️. TabPFN, an open tabular foundation model 🧠, reads your day like a row in a frost table and answers:

⏸️ Wait · 🚶 Walk · 🌿 Garden

Then it hands you three concrete steps and expects you to leave. The screen's job is under a minute. ⏱️


🚀 Live demo	plot-touch-grass.onrender.com
💻 Source	github.com/PranjalManhgaye/plot-touch-grass

🛠️ What I Built

Field	Value
🌿 Product	PLOT - outdoor action signal + checklist
🧠 Brain	TabPFN (`tabpfn-client`)
📊 Training data	30,684 real daily rows, 12 US cities, 2018-2024
🌦️ Weather	Open-Meteo ERA5 reanalysis + live forecast
☁️ Hosting	Render free tier
⚡ Stack	FastAPI, static UI, no chat layer

👥 Who it's for: anyone with a yard, balcony, or park nearby who overthinks the forecast and under-leaves the house.

🎬 Demo

🎥 Screen recording: [https://www.youtube.com/watch?v=0zisIS7GLjg]

🔗 Try it yourself: plot-touch-grass.onrender.com

⏳ First load on Render free tier may take ~30s while the service wakes up.

📸 Screenshots

🏔️ Hero - PLOT landing page

🌦️ Live weather - Open-Meteo fills the form

🎯 TabPFN signal - wait / walk / garden with confidence

⚙️ How it works (60 seconds)

flowchart LR
    subgraph YOU["You"]
        A["Pick city or use location"]
    end

    subgraph DATA["Open-Meteo"]
        B["Today's temps, rain, wind, soil"]
        C["92-day frost window"]
    end

    subgraph AI["TabPFN"]
        D["Classify: wait, walk, garden"]
        E["Class probabilities"]
    end

    subgraph OUT["Outside"]
        F["3-step checklist"]
        G["Phone down"]
    end

    A --> B
    A --> C
    B --> D
    C --> D
    D --> E
    E --> F
    F --> G

🗺️ User journey

sequenceDiagram
    participant U as User
    participant UI as PLOT UI
    participant OM as Open-Meteo
    participant T as TabPFN
    participant Y as Your yard

    U->>UI: Open site / pick Portland
    UI->>OM: GET /api/weather
    OM-->>UI: Real temps, frost lag, soil
    U->>UI: Get outdoor signal
    UI->>T: POST /api/check
    T-->>UI: garden, 83% confidence
    UI-->>U: Checklist (3 steps)
    U->>Y: Goes outside

🏗️ Architecture

graph TB
    subgraph Client["Browser"]
        WEB["Static UI / web"]
    end

    subgraph RenderHost["Render - plot-touch-grass"]
        API["FastAPI / backend/main.py"]
        ENG["OutdoorEngine / TabPFNClassifier"]
        CSV[("outdoor_history.csv - 30684 ERA5 rows")]
        WX["weather.py - Open-Meteo + fallback"]
    end

    subgraph External["Open services"]
        OM["Open-Meteo API"]
        TP["Prior Labs TabPFN"]
    end

    WEB -->|/api/weather| API
    WEB -->|/api/check| API
    API --> WX
    WX --> OM
    API --> ENG
    ENG --> CSV
    ENG --> TP

📈 The data story (real, not synthetic)

This is the part I cared about most. No fake temperatures. ✅

flowchart TD
    A["Open-Meteo Historical API - ERA5"] --> B["scripts/fetch_real_data.py"]
    B --> C["12 cities x 7 years of daily rows"]
    C --> D["Feature engineering"]
    D --> D1["days_since_last_frost from min temp at 0C"]
    D --> D2["soil_moisture, wind, daylight"]
    D --> E["Labels via backend/labels.py"]
    E --> F["outdoor_history.csv - 30684 rows"]
    F --> G["TabPFN fit and predict"]

🌎 Cities in the training set

City	Climate	Rows
🌲 Portland, OR	Cool marine	~2,557
🌧️ Seattle, WA	Marine west coast	~2,557
🏙️ Chicago, IL	Continental	~2,557
❄️ Minneapolis, MN	Upper Midwest	~2,557
⛰️ Denver, CO	High plains	~2,557
🗽 New York, NY	Humid subtropical	~2,557
🦞 Boston, MA	Coastal temperate	~2,557
🌳 Atlanta, GA	Warm humid	~2,557
🤠 Austin, TX	Warm	~2,557
☀️ Phoenix, AZ	Desert	~2,557
🌴 Miami, FL	Tropical humid	~2,557
🌊 Los Angeles, CA	Mediterranean	~2,557

📊 Label distribution (on real weather)

Action	Meaning	Share
🚶 walk	Good to be outside, not a planting day	~57%
⏸️ wait	Frost risk or bad conditions, prep indoors	~30%
🌿 garden	Frost window + workable soil, hands in dirt	~13%

Labels come from transparent horticulture heuristics in backend/labels.py (frost-free days, precip, wind bounds). TabPFN learns patterns in the table, not chat prose. I'm upfront about that. 📝

🔨 How I Built It

1️⃣ Real training pipeline

python scripts/fetch_real_data.py

Pulls daily ERA5 weather from Open-Meteo for every city above. Regeneratable. Documented in data/DATA_SOURCES.md.

2️⃣ TabPFN as the classifier

from tabpfn_client import TabPFNClassifier

model = TabPFNClassifier()
model.fit(X_train, y_train)           # real rows
proba = model.predict_proba(X_today)  # wait / walk / garden

Open-weight tabular model. Swap it without rewriting the product. 🔁

3️⃣ Live weather endpoint

GET /api/weather?lat=&lon= returns today's conditions + frost lag.

🌐 Primary: Open-Meteo forecast with past_days=92 for frost history
🛡️ Fallback: nearest city's real ERA5 row for today's calendar day (when API rate-limits on free hosting)

4️⃣ UI that stays out of the way

No chat bubble. No purple gradients. One brand (PLOT), one CTA, one result card. ✨

🏙️ City presets across 12 climates
📍 Use my location (browser geolocation)
🔄 Refresh live weather
✅ Checklist that ends at the door

5️⃣ Deployed on Render

render.yaml blueprint, health check at /api/healthz, TABPFN_TOKEN in env (never committed). 🔒

💻 Code

PranjalManhgaye / plot-touch-grass

PLOT - TabPFN outdoor go-signal for Hacktoberfest Touch Grass

PLOT

Leave the screen. Tend the ground.

PLOT is a Hacktoberfest Week 1 (Touch Grass) project. It uses Prior Labs TabPFN on real weather data (Open-Meteo ERA5) to classify today’s outdoor action as wait, walk, or garden — then gives a short checklist meant to end at your door.

Live demo

App: https://plot-touch-grass.onrender.com
Repo: https://github.com/PranjalManhgaye/plot-touch-grass

Pick a city or use your location → live weather loads → TabPFN predicts → go outside.

Real data (not synthetic)

What	Source
Training rows	30,684 daily observations, 12 US cities, 2018–2024
Weather	Open-Meteo ERA5 reanalysis
Frost lag	Computed from real daily min temps ≤ 0 °C
Labels	Transparent horticulture heuristics (`backend/labels.py`)
Live inputs	Open-Meteo forecast + 120-day archive

See data/DATA_SOURCES.md for full provenance.

Quick start

python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # set TABPFN_TOKEN
PYTHONPATH=. uvicorn backend.main:app --reload --host 127.0.0.1 --port

…

View on GitHub

plot/
├── backend/
│   ├── engine.py      # TabPFN + transparent baseline
│   ├── weather.py     # Open-Meteo live + ERA5 fallback
│   ├── labels.py      # Documented outdoor-action heuristics
│   └── main.py        # FastAPI routes
├── data/
│   └── outdoor_history.csv   # 30,684 real ERA5 rows
├── scripts/
│   └── fetch_real_data.py
└── web/               # Static UI

🔓 Why Does Open Innovation Matter?

Frost and planting decisions are tabular 📋. They belong in a spreadsheet, not a chat thread. 💬❌

🔒 Closed approach	🌱 Open approach (PLOT)
Garden API locks your coordinates to a vendor	Open-Meteo: free ERA5, no key
Chat model guesses from prose	TabPFN: probabilities over features
Can't explain why without prompt engineering	`garden: 83%, wait: 9%, walk: 8%`
Per-token cost at scale	Tabular fit once, predict cheaply
Model swap = rewrite the app	Swap TabPFN version, keep the CSV

🔐 Privacy: we don't store your location. Weather is fetched per request. Training table ships with the repo and is auditable.

🏕️ Offline path: the CSV + open model path means this could run on a laptop in the backcountry with cached rows. That's the Touch Grass future I want.

A closed chat API wouldn't know what to do with days_since_last_frost: 208. TabPFN was built for exactly that column. 🎯

🏆 What I'd tell judges


🌿 Theme fit	Screen is the shortest part; checklist ends at the door
🧠 Real open AI	TabPFN on 30k+ ERA5 rows, not a wrapper
📝 Honest labels	Heuristics documented, not pretending it's human survey data
🚀 Shipped	Live URL, GitHub, Render deploy
🚫 No chatbot cosplay	Tabular foundation model where it actually belongs

🤖 My Agent Session

Built with TabPFN docs + Open-Meteo API. Happy to answer questions in the comments. 💬

🎁 Prize Categories

🧠 Best Use of TabPFN - core classifier on real frost/weather tables
☁️ Best Use of Render - production deploy at plot-touch-grass.onrender.com

🔗 Links

Resource	URL
🚀 Live app	plot-touch-grass.onrender.com
💻 GitHub	PranjalManhgaye/plot-touch-grass
🌦️ Open-Meteo	open-meteo.com
🧠 TabPFN	docs.priorlabs.ai
☁️ Render	render.com

Thanks for reading. Now go touch some grass. 🌿👋

Top comments (11)

Tomasz • Oct 10

How often do you retrain the TabPFN model with new weather data to keep the predictions up‑to‑date?

Pranjal Manhgaye • Oct 10

TabPFN fits at startup on our ERA5 history CSV; today's weather row is fetched live from Open-Meteo per request. Full dataset refresh = re-run fetch_real_data.py + redeploy. No scheduled retrain yet; that's a solid v2 idea.
thanks