molecheck.info

A Doctor Built a "Skin Cancer Tinder" in 3 Hours with Gemini

Sungam, a dermatologist, posted on Show HN that he built molecheck.info using the free version of Gemini Pro 2.5 in roughly 2-3 hours. The entire app is a single file—HTML, JS, and CSS all in one—no backend, scores persisted via localStorage. That's the whole stack. His own words: "it's a very simple application but I would never have had the time to do it myself." This is the most honest articulation of vibe coding's value: it's not about building something complex, but about enabling someone with deep domain knowledge but no coding time to finally ship what's been in their head.

Swipe Left/Right Isn't a Gimmick—It's Clinical Triage

The interaction is dead simple: look at a skin lesion image, swipe left for "concerned," swipe right for "not concerned," with a third "I'm not sure" button. HN user k2xl tried it and marveled: "this game just proves to me how difficult your job is. I am basically getting 50%." But the more revealing feedback came from jjallen, who went from 50% accuracy to 85% quickly because "most of them are skin cancer and that was easy to learn." He suggested a 50:50 ratio of cancer to benign, and sungam agreed, adding that in reality most lesions are harmless—the real challenge is spotting the few that aren't.

Clinical Experience Is the Real Moat

HN user agnishom nailed it: "The main 'algorithm' to be implemented is very straightforward, and for the hard stuff, we have an expert." The code is essentially an image viewer with a scoring counter—any JS developer could build that. But which images to include, how to categorize them, when to offer the "unsure" option—that entire judgment framework comes from clinical experience. User jampekka raised a genuine concern: basal cell carcinomas look nearly identical to everyday rashes, and "my correct rate was under chance. This could be hypochondria inducing for many?" Sungam's reply showed a clinician's nuance: the key difference is temporal—rashes resolve, cancers don't; they slowly grow. A pure developer scraping dermatology image databases could amass hundreds of pictures but couldn't construct this cognitive framework.

Honest Limitations: A Trainer, Not a Diagnostic Tool

Several real issues surfaced in the thread. User aegypti drifted off after about 20 images—the app has no defined endpoint, and sungam committed to adding one. Leetrout found a Safari mobile bug causing image offset after submitting a guess; sungam admitted "no idea, I will look into this." The deeper concern is data distribution: alanfranz pointed out that real-world training would have a malignancy rate far below what the app shows—perhaps only 1% of lesions people worry about are actually cancerous. The current dataset over-represents cancer, which inflates accuracy and could fuel unnecessary anxiety. For the 429 points and 257 comments it earned, this is a genuinely useful learning tool—but nobody should mistake it for a replacement for professional diagnosis.