.png&w=256&q=75)
🤓 Latest Submissions
HemaVision Triage Agent
Anemia affects over a billion and a half people worldwide, and in Pakistan it's especially common among women and children. Diagnosing it usually means a blood test, which isn't always easy to get in rural or low-resource areas. There's a known non-invasive alternative: looking at the inner eyelid, since paleness there often signals low hemoglobin. The problem is, most AI tools built for this just give a number and call it a day, with no way to know if that number is trustworthy. We ran into this ourselves. While benchmarking a MobileNetV2 model on this exact task for a research paper, its accuracy swung by plus or minus 8 percentage points depending on how the data was split a single test could show 72% when the real, cross-validated number was closer to 62%. That gap is the reason HemaVision exists. Instead of hiding that instability behind one confident prediction, we built it into the product. Every image goes through 30 slightly-randomized passes through the model, and if those passes disagree too much, the app says so flagging the result as uncertain and recommending a repeat check, instead of quietly giving a shaky answer with false confidence. A language model (gpt-oss-20b, running on AMD's MI300X GPUs via Fireworks AI) then turns the numbers into something readable: a risk level, suggested tests, nutrition pointers, and how urgently to see a doctor. We tested it with a real phone photo, not a clean lab image, and it worked as hoped: since that photo differed from training data, it flagged its own uncertainty instead of guessing. To be upfront, this isn't a finished medical product. It's a screening aid, not a diagnosis, and the uncertainty threshold still needs testing on a bigger dataset. But that honesty is the point a tool that knows what it doesn't know is more useful, and safer, than one that always sounds sure of itself.
13 Jul 2026
