Household Guardian
A system that quietly watches over an elderly parent living alone — falls, unusual silence, phone scams — and sends their adult children one weekly message that usually says nothing happened.
I work on systems that act without a human reviewing each decision. Billing is where I do that professionally, and the consequences there are financial. This is the same problem with physical consequences, which is why it's interesting rather than a hobby.
The problem it exists for
Someone in their forties lives a few hundred kilometres from an ageing parent and worries daily. The available options are all bad. A panic button only works if it's worn and pressed, which is exactly what doesn't happen during a fall. A camera solves the technical problem and destroys the thing being protected, which is the parent's dignity in her own home. Several unconnected apps produce alerts nobody correlates.
Meanwhile voice cloning made phone scams substantially better, and older people are the preferred target.
The hard part
Telling a fall from everything that isn't one. A person going down and a dropped laundry basket look similar to a motion sensor. So do sitting down heavily, a dog jumping off a sofa, and a door swinging shut. The classifier isn't the hard part — the hard part is that both failure directions are expensive in different ways, and they trade against each other.
Miss a real fall and someone lies on a bathroom floor until morning, which is roughly what determines whether a fall is survivable. Raise a false alarm often enough and the family stops answering. That second failure is the one I take more seriously, because it's silent: the system keeps working, the alerts keep arriving, and nobody is reading them any more. It's the same failure as an escalation queue that humans have learned to approve without looking.
Deciding what counts as normal, per household. There is no population baseline worth using. Someone who gets up twice a night is unremarkable if she always has, and a signal if she didn't last month. The system has to learn a specific person's ordinary before it can flag a departure from it, which means it's least reliable exactly when it's newest.
Sensing without recording. The constraint is that no camera and no microphone exist anywhere in the design — not disabled, not opt-in, not encrypted. Not present. Detection uses radio reflection, the same principle as a car's parking sensor: it can tell that someone is upright, moving, or on the floor, and it cannot tell what they look like or what they said. Nothing is captured, so nothing can be leaked, subpoenaed, or accidentally shown to the wrong person. That constraint costs real accuracy and I am not willing to trade it, because a system that protects someone by surveilling them has misunderstood what it was asked to protect.
Deciding what happens when it's unsure. Uncertainty is the normal case, not the exception. A system like this needs three distinct responses — act now, mention it later, ignore it — and most of the engineering is in the boundaries between them. Getting up more often at night is not an emergency and not nothing; it belongs in a weekly note as something worth raising with a doctor, not in a 2am phone call.
What it deliberately does not do
It does not call an ambulance. It does not diagnose. It does not decide that someone should stop living alone. Every one of those is a decision with consequences a family should own, and routing them through software would move accountability somewhere it does not belong. The system's job ends at telling a human something they would otherwise not have known, in time to act on it.
It also does not talk to the parent, which was a deliberate reversal. The obvious design has the system ask "are you alright?" after a suspected fall. That makes it a device demanding a response from someone who may be unconscious, and turns a silence into ambiguous evidence.
Where this is honest about being unfinished
The scam-call detection is the least mature part. Blocking a known-bad number is straightforward; recognising a cloned voice mid-call, fast enough and reliably enough to intervene without cutting off a genuine grandchild, is not solved. The current answer is deliberately low-tech — a family safe word, on a card, next to the phone — because it works today and doesn't depend on the model being right.
The other open question is what happens the first time the system is confidently wrong about something that mattered. I have opinions about the audit trail and none of them have survived that event yet, because it hasn't happened.