Evidence
We will show results as soon as we can substantiate them. Internal test scores are not the same as a real home. We will not compete by claiming higher percentages than other vendors.
Current stage: first pilot started on 5 October 2026 in Ilona's home; room for four more in this pilot. No facility deployments yet. The detector works on our own test material and on staged falls in the founder's home. From the pilot we will publish false alerts, uptime, and staged detection in situ. Real falls are rare; if one happens, we will report it as a case, not as a rate.
What can be said when
The order below keeps future goals from being read as results.
Architecture facts, once implemented and technically verified
These describe how the system is put together. They say nothing about real-world outcomes:
- AI runs locally
- no live camera feed
- no audio recorded
- normal imagery is not transmitted
- rolling buffer duration
- event-image policy
- camera-vendor cloud not required
- dedicated network / cellular architecture
- power-loss notification behavior
- heartbeat behavior
After controlled benchmarking, labelled “internal test”
- sensitivity on a named test set
- false-positive performance on a named test set
- model latency
- compute capacity
Every such number comes with its dataset size, the split between fall and non-fall clips, whether the rooms appeared in training, how much was synthetic, which camera viewpoints were used, and confidence intervals where they make sense.
If an internal held-out set looks extremely strong, we will still call it internal held-out test performance. It is not real-world performance.
After home pilots
- false alerts per home per month
- staged detection in the live homes (labelled as staged)
- any unstaged fall, reported as a case note
- night / IR behavior
- camera coverage
- alert delivery time
- system uptime
- detection rate on staged falls
- latency and alert delivery
What we will measure (and why)
| Metric | Why it matters |
|---|---|
| Fall sensitivity (staged / recorded) | Share of staged or recorded falls that are detected; labelled as such |
| Unstaged falls | Reported as case notes when they occur; too rare for a rate in a small pilot |
| False alerts per resident-month | What families and staff actually experience |
| Median alert latency | How quickly the event is surfaced |
| 95th-percentile latency | Worst normal-case delay |
| Slow-descent sensitivity | Hard, realistic fall category |
| Night / IR performance | Homes are dark |
| Occluded-fall performance | Furniture is unavoidable |
| Sitting / lying-on-floor false alerts | Common real-world confusion |
| Camera uptime | Coverage reliability |
| AI-node uptime | System reliability |
| External alert delivery | End-to-end, not just the model |
What we will not headline
- “100% accurate” or “zero falls missed,” even if an internal set is clean. Real homes are harder, and those slogans are already used elsewhere.
- ilonai does not claim to reduce falls or their consequences.
- “We prevent falls.” What the product does is detect a possible fall and alert the people you choose.
- Unsourced ROI percentages, invented pilot conversion rates, or certification dates that were never met.
Context
Falls when alone are common in older people. Swiss figures come from the BFU, September 2026: more than 1,700 fall deaths a year in Switzerland, 95% of them aged 65 or over.
Reliability tests we intend to run and then publish
With a working home installation in place, we will run and then publish:
- cut the WAN / cellular path
- cut mains power
- unplug a camera
- reboot the local AI computer
- restore everything
A reliability write-up will appear here once those tests have been done, and not before.