Stopping on its own once both readings settle is a nice touch. One design choice I'd consider for the tips: pick them from a short fixed list that someone has reviewed, rather than letting the LLM write free text from a single reading. Camera readings move with lighting and movement, and a fixed list keeps a jumpy number from ever turning into a confident sounding health message. Do you show how stable the reading was alongside the result?
Thanks @gautamtalksdev . I actually implemented this flow already. The camera session automatically stops once the readings become stable, and the LLM is only called when the user explicitly clicks the “Tips” button. So the LLM doesn’t continuously interpret the live readings or generate advice from every fluctuation. The next thing I’m considering is making the stability/signal-quality information more visible alongside the final result.
Nice, stopping the camera once readings stabilise and calling the LLM only on the Tips click is exactly the right shape. Showing signal quality to the user sounds like a good next step, people trust a reading more when they can see how steady it was.
Using rPPG and optical micro-fluctuations for contactless physiological monitoring is a fascinating frontier. The primary hurdle often boils down to ambient lighting variance and compression artifacts from commodity webcams, but modern vision models are closing that gap remarkably fast.
Nice job, I really like how you present this solution. You truly convey confidence about the application and how it handles the data, not storing anything sensitive.
My question is: have you been able to confirm the signal's reading accuracy? Like, comparing it to analog readings to determine how well the camera can read these signals. I feel kind of skeptical about this aspect.
Thanks. That’s a fair concern. I’ve checked the readings during development with my own measurements, and the signal capture has been working as expected. I still want to do a more thorough validation by comparing the readings with reference measurements, especially across different conditions and abnormal cases. I’m planning to do that validation over the weekend.
Stopping the measurement on its own once both readings settle, and only calling the LLM when someone clicks Get health tips, is a sensible shape. The line I liked most was that the AI only ever receives the two numbers, never who you are. Gautam's suggestion of picking tips from a short reviewed list is worth taking too, since a camera reading moves with light and movement.
We ran into the same "two numbers in, a lot of meaning out" pattern at Ascendra Ventures when we rebuilt the experience for Actofit body-composition scales after the official app stopped returning the full readout. The scale only sends weight and bioelectrical impedance over Bluetooth. Body fat, BMI and skeletal lean mass are all calculated from those two numbers and the user's profile with the openScale/Xiaomi body-composition model, in the browser over Web Bluetooth with nothing to install. That split between measured and derived numbers is something I think every health app should show plainly, and your typical, low and high bands are a good step in that direction.
When you validate against reference readings, will you also show users how steady each reading was, as you mentioned in the comments?
Thanks :) And yes, privacy is especially important in medical technology, so I wanted to make sure patient information is never exposed .
Right now, the measurement automatically stops only after the readings become stable. I’m also planning to show the stability/signal-quality information alongside the final readings, rather than showing only the two numbers.
We're a place where coders share, stay up-to-date and grow their careers.
Stopping on its own once both readings settle is a nice touch. One design choice I'd consider for the tips: pick them from a short fixed list that someone has reviewed, rather than letting the LLM write free text from a single reading. Camera readings move with lighting and movement, and a fixed list keeps a jumpy number from ever turning into a confident sounding health message. Do you show how stable the reading was alongside the result?
Thanks @gautamtalksdev . I actually implemented this flow already. The camera session automatically stops once the readings become stable, and the LLM is only called when the user explicitly clicks the “Tips” button. So the LLM doesn’t continuously interpret the live readings or generate advice from every fluctuation. The next thing I’m considering is making the stability/signal-quality information more visible alongside the final result.
Nice, stopping the camera once readings stabilise and calling the LLM only on the Tips click is exactly the right shape. Showing signal quality to the user sounds like a good next step, people trust a reading more when they can see how steady it was.
Yes !!
Using rPPG and optical micro-fluctuations for contactless physiological monitoring is a fascinating frontier. The primary hurdle often boils down to ambient lighting variance and compression artifacts from commodity webcams, but modern vision models are closing that gap remarkably fast.
Absolutely. That’s one of the interesting challenges I'm exploring :)
Nice job, I really like how you present this solution. You truly convey confidence about the application and how it handles the data, not storing anything sensitive.
My question is: have you been able to confirm the signal's reading accuracy? Like, comparing it to analog readings to determine how well the camera can read these signals. I feel kind of skeptical about this aspect.
Thanks. That’s a fair concern. I’ve checked the readings during development with my own measurements, and the signal capture has been working as expected. I still want to do a more thorough validation by comparing the readings with reference measurements, especially across different conditions and abnormal cases. I’m planning to do that validation over the weekend.
Stopping the measurement on its own once both readings settle, and only calling the LLM when someone clicks Get health tips, is a sensible shape. The line I liked most was that the AI only ever receives the two numbers, never who you are. Gautam's suggestion of picking tips from a short reviewed list is worth taking too, since a camera reading moves with light and movement.
We ran into the same "two numbers in, a lot of meaning out" pattern at Ascendra Ventures when we rebuilt the experience for Actofit body-composition scales after the official app stopped returning the full readout. The scale only sends weight and bioelectrical impedance over Bluetooth. Body fat, BMI and skeletal lean mass are all calculated from those two numbers and the user's profile with the openScale/Xiaomi body-composition model, in the browser over Web Bluetooth with nothing to install. That split between measured and derived numbers is something I think every health app should show plainly, and your typical, low and high bands are a good step in that direction.
When you validate against reference readings, will you also show users how steady each reading was, as you mentioned in the comments?
Thanks :) And yes, privacy is especially important in medical technology, so I wanted to make sure patient information is never exposed .
Right now, the measurement automatically stops only after the readings become stable. I’m also planning to show the stability/signal-quality information alongside the final readings, rather than showing only the two numbers.