I pulled up the sensor data from my parents' home last week. Not because something was wrong. Just to look.

The timestamps scrolled by — activity, quiet, activity, quiet — and somewhere in there I had a small, weird moment. I knew those patterns. I recognized them. That cluster of motion at 6am was my mom making coffee. The gap in the afternoon was my grandmother's nap. The late-night blip was, almost certainly, the bathroom.

The data was starting to look like a person.

Most of the last two weeks was spent building the algorithm that learns this. The core problem: population averages don't work. My grandmother has her own rhythm. The system needs to learn her specifically — not what elderly people typically do, but what she actually does.

I won't go deep on the math. But the thing I keep coming back to is a small decision I made about memory. The system only uses the last 14 days. Older patterns get discarded.

That felt almost philosophical when I was building it. The system should know recent her. Who she was six months ago matters too and that's actually the long game — but for now, two weeks is the right window to start.

For the first three days in a new home, it doesn't alert at all. It just watches. It earns the right to speak.

On the side, I've been building some agentic pipelines — essentially, code that reviews itself, catches its own bugs, and helps me move faster without introducing new problems every time I touch something. It's the kind of infrastructure work that isn't visible in the product but is the difference between a codebase I trust and one I'm afraid of. (Especially since Claude is writing all my code 🫥)

Less glamorous than the algorithm. Completely necessary.

I also want to give a proper shoutout to Anagha.

She's a UX researcher, currently in school, and she has spent the last few weeks helping me with something I was genuinely not equipped to do alone — user interviews, survey design, synthesizing what people actually said vs. what I wanted to hear. She built wireframes. She pushed back on my assumptions, challenged me. She did all of this while juggling her own coursework.

The kind of help you can't put a title on. She just showed up and cared about it.

Anagha, thank you. Seriously 💙

The algorithm is deployed. Two pilot homes generating real baselines. False positives still exist — I know where to look now, which is progress.

Next step: two or three more families. Starting with my in-laws and my husband's grandparents — they said yes before they fully understood what they were agreeing to.

There's something I didn't expect about this phase. I thought building the algorithm would feel like an engineering problem. It mostly felt like paying attention.

Figuring out the difference between waking up and a bathroom trip. Deciding how long "normal" should last. Teaching the system to be quiet for the first three days.

It's starting to know my grandmother. I find that either beautiful or slightly terrifying, and I haven't fully decided which.

See you in two weeks

— Shwetha