@_flixmd profile picture
Flix @_flixmd

Colorectal surgeon 🏥 | Building software between surgeries 💻 | Barcelona 📍 | Founder of @trialinx & @relaylit

Joined March 2026
  • Tweets 945
  • Following 224
  • Followers 45
Photos and videos
Flix@_flixmd
Jun 13
Clinical AI needs boring fallback plans. If model access changes overnight and the work stops, the risk was never just accuracy.
6
Flix@_flixmd
Jun 12
High medical AI usage isn't the same as implementation. The hard part is where the answer goes next.
18
Flix@_flixmd
Jun 12
If an AI agent can run for days in clinical work, the win isn't endurance. It's knowing exactly when it should stop.
19
Flix@_flixmd
Jun 11
A medical AI benchmark isn't a product. The product is the version clinicians can actually use.
16
Flix@_flixmd
Jun 11
A health app that spots patterns but can't hand them to the clinic is just a nicer diary.
13
Flix@_flixmd
Jun 10
Guardrails matter. But if medical AI blocks ordinary research, people will route around it. That's not safety.
19
Flix@_flixmd
Jun 9
Subagents in healthcare can't be little black boxes passing notes to each other. If the trail breaks, the clinician becomes the debugger.
14
Flix@_flixmd
Jun 9
Clinical trial software doesn't break in the demo. It breaks when old data, permissions, audit trails, and the forbidden spreadsheet all have to move together.
11
Flix@_flixmd
Jun 8
Medical AI can keep the data private and still dump the work back on the clinician. Privacy is necessary; workflow is the product.
11
Flix@_flixmd
Jun 8
Healthcare distribution is workflow trust. If a tool lives outside the chart, inbox, or study record, even a better answer has to beg for adoption.
10
Flix@_flixmd
Jun 7
Most clinical AI demos ask: can it answer? The better question is: what work should this answer create, if any?
13
Flix@_flixmd
Jun 1
Healthcare doesn't have a signal shortage anymore. Sleep scores, home scans, blood tests. The bottleneck is knowing which signals deserve action and which are just noise.
13
Flix@_flixmd
May 31
Early cancer detection doesn't end uncertainty. It moves uncertainty upstream, where the hard part is follow-up: explain the signal, avoid false certainty, own the next step.
15
Flix@_flixmd
May 22
Medical AI is getting better interfaces. Voice, chart context, fewer copy-pastes. Good. But the product still has to know when silence is safer than an answer.
19
Flix@_flixmd
May 21
Medical AI can't hide behind "clinician sign-off." If the system made a dozen quiet decisions first, the safety layer is the visible stops, not the final signature.
20
Flix@_flixmd
May 21
The next medical AI fight won't be "is the answer good?" It'll be "who let it become part of the chart?"
19
Flix@_flixmd
May 20
A medical AI demo can win the exam room and still fail the patient. If follow-up dies in an inbox, the product didn't automate care. It made a cleaner note.
39
Flix@_flixmd
May 20
The dangerous medical AI citation is not always the fake one. Sometimes the paper is real, the quote is real, and the claim still doesn't follow.
14
Flix@_flixmd
May 19
Healthcare AI doesn't need to cosplay as a doctor. It needs to know when it changed the work, when it's out of bounds, and when to hand it back.
18
Flix@_flixmd
May 19
AI-scribe evals should not stop at whether the note sounds right. The real failure is propagation: a wrong sentence becoming a med list, follow-up task, billing code, or research field. Score where the error can travel, who reviews it, and how it gets reversed.
1
23
Load more