This is the first entry in a running log about my work on Synheart AI’s research team. I want a record of the work while it’s happening — decisions look obvious in retrospect and never are at the time. One entry a week, roughly.
The problem space
Synheart works on human state recognition from kinematics — reading what a person’s movement says about them. The signals we care about split into a few families:
- Activity states — what the person is doing (walking, sitting, climbing stairs)
- Postural states — how they’re holding themselves while doing it
- Locomotion — the mechanics of how they move through space
- Movement regularities — the patterns and rhythms that make one person’s movement different from another’s
The long-term direction is folding all of these into an HSI (Human State Inference) model — one system that takes motion signals in and produces a coherent picture of the person’s state, rather than four separate classifiers that disagree with each other.
What week one actually looked like
Research week one is unglamorous in exactly the way I expected:
- Reading. A stack of papers on activity recognition and postural analysis, each of which I summarize for the team — writing the summary is how I find out whether I understood the paper.
- Data. Working on dataset collection for the states we want to model. The quality of everything downstream is capped by decisions made here, which is a familiar lesson from RAG work: the model can’t infer what the data never captured.
- First experiments. Getting baseline models running so we have numbers to beat, not vibes to argue about.
A note on why I’m logging this
My blog posts so far have been retrospectives — the compiler story, the payments story — written after the ending was known. This series is the opposite: notes from inside the work, before I know how it turns out. If you’re reading this as a hiring manager or a fellow researcher, that’s the point — this is what my week actually contains.
Next entry: baselines, and whatever the data collection breaks first.