I am a plant serial killer. I buy plants, place them around my house, and they die. Too much water. Too little water. Sun. Shade. If there is a narrow band of conditions in which a plant survives contact with me, I have never found it.
The flowers by my front door were the clearest case. They died, I replaced them, they died again. After a while, I stopped treating it as a problem to solve and started treating it as an ongoing expense.
So I asked Mira, my local AI assistant that runs on my computer, to remind me when it was time to buy the next lot. I was not asking her to fix anything. I was asking her to schedule my next anticipated failure.
Then she asked what kind of flowers they were.
The wind problem
I did not know what kind of plant I was trying to force onto my front door patio, so I sent her a picture. She came back and told me the area was too windy for that particular plant.
Wind. In all the years I had been drowning, dehydrating, and scorching things on that doorstep, it had never occurred to me that moving air was a variable. Great, one more way to kill a flower.
I asked her what I should buy instead. She gave me a list of candidates with Latin names. I realized this was not going to be helpful. If I wanted the perfect plants for my specific house, I would need to give her more input.
How Mira works
Mira is built to run locally on Ollama, using a 12B model[reference:40]. The entire concept is to give my laptop the ability to think and act on its own without some giant corporation consuming every piece of information I type[reference:41]. She can even write her own code and create tools to perform specific jobs. But that code does not touch my disk until I approve it. I wanted to stay in the loop rather than wake up to a machine that had rearranged my life.
We worked it out together. I explained ideas. She explained what made sense and what did not.
Finding the real data
Her first suggestion was that amateur weather stations exist, that many publish their readings openly, and that there was likely one near me. I told her to find out. She built herself a small tool to search for local stations. There was one a few houses away from mine[reference:42]. Not a municipal station on the edge of town averaging conditions across the region. This one was on the street where I walk my dog three times a day. I never noticed it.
She pulled its history and worked out the numbers that actually apply to my address: average rainfall, wind speed, cloud cover[reference:43]. I had been buying plants based on what looked nice. Now I had hard data.
I told her to look at my house on Google Earth from above and work out the sun's path across it over the year, then estimate how much each side receives[reference:44].
That is when my house stopped being one place and became four different time zones. One side is bright and dry for most of the day. One is in shade so consistently that it may as well be a different country. Another takes the wind. One is fine, apparently. Every plant I had ever bought had been chosen for "my house," as though a house was a single set of conditions. I had been running a four-climate operation and blindly shopping for one[reference:45].
Mira matched flowers to each side of the house. I went to my local plant store, pointed at my phone, and said "This."
What it cost
I want to be careful not to oversell this. It is only one season. Mira also reminds me to check on them when it has not rained in a while. But I know exactly what my hit rate was before, and it was zero[reference:46].
The interesting thing is not that a machine gave me better advice than I could find myself. It is what it cost. Mira runs on my own hardware, on a 12B model. She has an escalation path to a bigger cloud model for anything she cannot handle, but I burned through $15 of API credit in roughly one session. The garden was done entirely by the small local one that was never supposed to be the clever half[reference:47].
For about 10 years, I let companies sell me smart home products. What I got was lights that come on when I walk into a room, and even that did not always work[reference:48]. I have come to think the category sold me eyes and ears, but never sold me a brain. A motion sensor is like a reflex. It fires, something happens, and nothing anywhere in the house has had a moment's thought about it[reference:49].
That is not smart. Ten years of that, and the most useful thing my home has ever done for me came out of a conversation, needed no hub, no bulb, no plug, and no sensor, and involved a stranger's weather station and a satellite photo of my own roof[reference:50].
Now, you might say this is not a smart home. Mira is not wired into anything. She cannot turn on a light, lock a door, or open my blinds. By the standards of the aisle I used to shop in, my house is exactly as dumb as it was ten years ago[reference:51].
Except every product in that aisle is sold on doing something for me, and not one of them has ever known the first thing about where I live. Automation was always the easy half. What I wanted was something that understood the place, and that was never on the shelf[reference:52].
The flowers outside are thriving. The ones inside are still dying because I am secretly still a plant serial killer. But at least my neighbors do not know that.