PredictCare
Type:
ML-backed full-stack product
Stack:
Next.js, Flask, scikit-learn, XGBoost
Result:
95%+ model prediction accuracy
A disease-likelihood tool with a Next.js front end and a Flask service behind it, built so that a non-technical user can get an assessment without being handed a probability they cannot interpret. The headline number is 95%+ prediction accuracy, but the number is the output of the process, not the point of it.
The process was model selection done properly. Logistic regression, random forest and XGBoost were applied to build the classifier, and the three highest-performing models were chosen from more than fifteen evaluated. Most of the work in applied machine learning is the fourteen you discard.
The product side matters too: a Suggestion Page introduced the first feedback loop into the application and captured 50+ user suggestions — which turns a model demo into something with a channel back from the people using it.
3 models, chosen from 15+
Logistic regression, random forest and XGBoost evaluated against each other, with the three best integrated for accuracy rather than picked for familiarity.
Next.js over Flask
A front end built for people without a statistics background, served by a Python service doing the actual inference.
50+ suggestions captured
The Suggestion Page added the product’s first feedback system, turning a classifier into something users could talk back to.