What the "AI in healthcare" conversation keeps getting wrong (from someone who built one)---Everyone building AI health too
Everyone building AI health tools is copying the same architecture:
user uploads image → cloud API → result returned.
Looks clean in a demo.
Falls apart economically at scale.
I went a different direction with my skin health app.
TFLite model runs entirely on-device:
No per-call inference cost
No latency
No privacy concern about images leaving the phone — because they don't
A few things I figured out the hard way:
On-device inference changes your unit economics permanently
When the model runs locally, your cost per scan is $0 — regardless of volume.
That’s not a small detail.
It’s the difference between a viable consumer product and one that gets squeezed by its own growth.
Privacy-by-architecture beats privacy-by-policy
“We comply with GDPR” is a legal statement.
“Your scan runs on your phone and never leaves it” is a product statement.
Users respond very differently to the second one.
The AI model is not the moat
Anyone can fine-tune a skin analysis model.
What actually takes time:
Two-sided dermatologist marketplace
Patient data sharing
Async consultations
Subscription tiers
Admin tooling
The model is maybe 10% of the real product surface.
$0/month infrastructure is a strategic choice, not a limitation
Full stack running on:
Google Sheets
Apps Script
Cloudinary
Netlify
The right question isn’t “can it handle 10M users?”
It’s “can it reach the point where you can afford to migrate?”
Yes. Easily.
Curious if anyone else building health or wellness AI has hit similar decisions — especially around on-device vs cloud inference tradeoffs.
What pushed you one way or the other?
(If you want to see the actual build — Flutter, full derm booking marketplace, $0/mo infra — I posted about it separately. Happy to answer architecture questions here too.)
