Future work

Google Flood Hub already forecasts every major river in India seven days out. Its own line is "move people, not water". So the gap is not warning, it is deciding when a reservoir should release. I built a simple reservoir-operation model and it gave an asymmetric answer: releasing water early helps in long, slow floods and makes things worse in short cloudbursts. That turns the whole problem into telling the two kinds of event apart early enough to act differently.

Next step: write to the academics whose group has studied exactly that distinction, leading with a question only they can answer rather than a request for data. The draft is ready. No emails sent yet.

The task design is settled, including running the same prompt once and forty times to separate luck from ability. The bottleneck is traces, so collecting them comes before anything else.

I assumed the clever part would be the connector that detects which module is plugged in. It is not: anyone can copy that in a few days. What is hard to copy is the data the tutor collects, student code plus compiler errors plus what the hardware actually did, and the promise that it all transfers. So the first thing built is the tutor's evaluation harness, with a deliberately bad tutor to prove the harness can catch one. Curriculum is written and mapped to MYP, Cambridge and IGCSE. It needs a teacher to read it next.

My first round of ideas all quietly assumed a display, and every one of them died when I checked the hardware. The build plan now starts with the boring shared part, the capture loop, the voice, a quiet mode and a rule to say things once, because that is where you find out whether wearing it would be annoying. I will only buy the glasses if that test passes.

The problem in one line: a 2 to 3 percent change in blood oxygen is both the signal I am looking for and the size of the sensor's own error. One person's bad night is noise; the same shift across 200 dose changes is signal. So the first step needs no hardware at all: look for the effect in the All of Us dataset, which is free and already collected. The moment the output would change someone else's dose it becomes a medical device, and I know where that line is before building towards it.

I expected the software to be the hard part. It is the easy part: background cleanup, AI try-on and pushing to a shop platform are a few weeks. The mechanics are months, because clothes swing, fold and catch. Several conveyor designs died on that. The carousel won because the garment touches nothing but its own hanger. Next step is to build it without a motor, turn it by hand, and add the motor later as a one-evening upgrade. Hand-turned, it already gives about 95 percent of the value.