![Cover image for [Day 10] Building my own personal weather officer AI, and teaching it my body's sense of cold over the next 100 days](https://media2.dev.to/dynamic/image/width=1000,height=420,fit=cover,gravity=auto,format=auto/https%3A%2F%2Fraw.githubusercontent.com%2FSAETAG%2Fdgx-100-experiments-images%2Fmain%2Fday10-weather-officer%2F01-swing.png)
Intro
Day 10!
This time I'm starting a longer-running experiment. Meet the "weather officer AI" ā a bot that texts me every morning saying "wear this today." The plan is to build a weather assistant that's tuned to me.
What I'm building today is just v0.1 (the very first version). From here through Day 100, I'll keep teaching it "too cold / just right / too warm" every morning, so it gradually learns my preferences. The experiment is: how smart does it get after 100 days?
What I used: my home AI machine (DGX Spark) + free weather data + a phone messaging app (Telegram)
Today's task
What I wanted
I live somewhere with a big daily temperature swing, and "what do I wear today?" is a small but real daily headache. Weather apps tell you the temperature, but whether I feel cold is a different question.
So the starting point was: can I build a clothing AI that's tuned to "how I feel," not just "the temperature"?
Approach
I kept the design dead simple.
- Every morning at 7, automatically fetch today's weather
- Decide "this morning's outfit" from the apparent temperature and push it to my phone
- I just tap back "cold / just right / warm"
- As these feelings pile up, the AI learns "this person runs cold" and corrects its suggestions
The goal this time
Not a "perfect forecast AI," but a "routine I can actually keep up every day." The smarts get grown over the next 100 days. Today is just laying the rails.
š How much does the temperature actually move in a day?
Before building anything, I pulled a week of apparent temperatures for where I live and graphed it.
A beautiful zigzag. Every day repeats "cold in the morning ā way up by midday ā down again at night." The average daily swing is 13Ā°C, and on the biggest day it moved 20Ā°C.
So I narrowed the suggestion down to "one outfit, at 7 a.m., matched to the apparent temperature at that hour." I record just once in the morning too. To keep something up for 100 days, simplicity matters most.
The AI doesn't learn the temperature swing itself. But by deciding to "focus on the morning," the suggestion and the feedback line up in time, so later I can cleanly check "was the morning suggestion right?"
š§ How it works (the morning loop)
The finished weather officer runs on this loop.
At 7 a.m. it grabs the weather, decides the outfit, and pushes a notification. I tap back my feeling, and that gets recorded. Once those records pile up, step 5 ā learning "you run cold / warm" ā kicks in, and the suggestions gradually become mine.
Here's what the actual notification looks like.
āļø Weather Officer AI ā good morning
š This morning's outfit: long sleeves
This morning feels like: 13Ā°C (highs up to 20Ā°C today)
Rain: 4% / Wind: 13 km/h
How does it feel this morning? ā tap to tell me
[š„¶ cold] [š just right] [š„µ warm]
It suggests one outfit, but adds "highs up to 20Ā°C today" so I can decide whether to throw on a layer myself. Tapping a button changes it to "ā recorded," and the feeling is saved to my home AI.
The notifications go through the messaging app I already use (Telegram).
š ļø The details
Below are the specifics.
The weather data
Weather comes from Open-Meteo, a free weather API. No API key needed, historical data available, and commercial use is OK (CC BY 4.0) ā very generous.
I mostly use "apparent temperature" ā not the raw air temperature, but a number adjusted for wind and humidity to reflect how it actually feels, which is better for deciding what to wear. I take the average apparent temperature from 7ā9 a.m. as "this morning's feel." The coordinates stay only on my home machine; I don't write the specific location in the article or the code.
The clothing rule
A plain rule that splits apparent temperature into 7 bands, each mapped to an outfit (e.g. 13ā20Ā°C ā long sleeves, 20ā26Ā°C ā short sleeves).
There's one "personal offset" number baked in. It's zero for now (v0.1). Going forward, if "cold" keeps coming back, I'll push the offset negative so the same temperature suggests warmer clothes ā growing it from the feedback.
The notification and buttons
The notification uses a Telegram "Bot" (a thing that sends messages automatically), built with the python-telegram-bot library to:
- send a message at a fixed time every morning at 7
- attach three buttons below it and record which one is pressed
This bot sits waiting inside my home AI machine and fires in the morning.
The shape of the feeling log (the 100-day foundation)
Each record is one line per day, with these fields:
- date
- that day's forecast (morning/daytime apparent temperature, wind, rain chance)
- the outfit the AI suggested
- my feeling (cold / just right / warm)
Kept in this consistent shape, I can later graph "monthly hit rate" or "my personal bias." I'll report this trend in the Day 21 / 36 / 54 / 74 / 87 check-ins.
Keeping it running
So the 7 a.m. notification reliably fires, I set the bot to launch automatically when the machine boots (systemd). Even after a reboot it comes back on its own, and the morning notification keeps going.
The feeling log and the notification settings (the Telegram token, etc.) are all stored only on my home machine ā none of it goes anywhere external like GitHub.
The 100-day growth plan
I'll keep growing this weather officer across the series ā it'll pop up here and there.
| Milestone | What happens |
|---|---|
| Day 10 (now) | v0.1 done; the recording loop starts turning |
| Day 21 | First "my personal bias" report from 11 days of data |
| Day 36 | Graph the monthly hit rate and take a look |
| Day 54 / 74 / 87 | Mid-reviews: seasonal changes in feel, how the correction works |
| Day 100 | The 100-day accuracy trend and the finished version |
Just tap a button every morning. I'm curious how "mine" it'll feel after 100 days.
Next time: Day 11
Next time it's a hard pivot back to cats š± I'll convert photos of my cat into picture-book, anime, and photorealistic styles ā the theme being whether I can keep "that's-our-cat-ness" while changing only the style.
Top comments (1)
this is a cool approach to personalizing your daily decisions based on how you feel. tuning an AI to learn your preferences over time sounds like a fun challenge. if you're ever interested in building an app around your ideas, moonshift lets you get a full next.js + postgres + auth app deployed in about 7 minutes, and you own the code. happy to help you get a free run on it if you're up for it.