MathFi.ai
MathFi provides deterministic machine-learning models that detect rare, high-impact events like strokes, heart failures, and fraud. Users upload labeled data in the morning and can call a CPU-only model by the afternoon, with every prediction reproducible and auditable. The platform includes a Feature Refinery tool for automated column engineering and runs hosted, in the customer's cloud, or in a private environment.
- Artificial Intelligence
- Healthcare Technology
- Data & Analytics
- Software Only
Funding
Founders
Founder details are not available yet.
Product
Problem
Standard machine-learning models often fail to detect rare, high-impact events because they optimize for overall accuracy, which rewards predicting the common class almost every time. This leaves critical cases—such as stroke patients, failing signals, or fraudulent claims—undetected, and the models themselves are often opaque, making predictions difficult to audit or defend.
Solution
MathFi provides deterministic machine-learning models that specialize in detecting rare, high-impact events such as strokes, heart failures, and fraud. Users upload labeled data in the morning and can call a CPU-only model by the afternoon, with every prediction reproducible and auditable. The service runs as a hosted solution, in the customer's cloud, or in a private environment, and includes a Feature Refinery tool that automates column engineering for data teams. The platform measures performance on the minority class—the rare cases that matter—rather than headline accuracy, and benchmarks show higher minority-class F1 scores than standard libraries like XGBoost and LightGBM across multiple public datasets.
Target Audience
Primary users are data science teams and analysts at organizations in healthcare, finance, insurance, and manufacturing that need reliable detection of rare, high-stakes events and auditable model outputs.
Features
- Deterministic models: same data in, same decision out, with every prediction replayable and defensible to an auditor
- CPU-only inference: no GPUs, clusters, or queues required; models are callable by the afternoon after uploading labeled data in the morning
- Deployment flexibility: hosted service, customer's own cloud, or private cloud for data that cannot leave the environment
- Feature Refinery: automated column engineering that handles the data preparation work scientists often lack time for
- Data Visualiser: profiles every column on upload, showing types, missing values, distributions, and target class balance before training
- Minority-class performance focus: measured on recall and F1 for the rare class, with documented benchmarks against five standard libraries on seven public datasets