Feature Refinery
Works out which of your columns actually carry signal, and drops the rest. Weeks of feature
selection becomes one run.
Model Crucible
Trains every algorithm family against your data at once, scores them on data they never saw,
and keeps the best three. You choose which one predicts.
Four minutes, end to end, on a public credit-risk dataset. It starts muted and the on-screen
captions carry it, so unmute only if you want the sound. The same run is written up step by step in
German Credit.
Where it earns its place
MathFi.ai works on any labelled, tabular data. It is strongest where the answer that costs you money is rare and easy to miss: the fraud that clears, the machine that fails, the claim that should have been challenged, the good customer you wrongly declined. That is the case ordinary methods handle worst. Gradient-boosted trees optimise for being right most of the time, and on a dataset that is 3% positive, “right most of the time” means predicting negative and looking accurate. MathFi.ai gives the rare class its own structure, so precision and recall hold up where they matter. Seven benchmarked use cases, and whether yours looks like them →What makes it different
Two settings, not a hundred. A tool like XGBoost exposes well over a hundred knobs. A MathFi.ai model has two, and the platform works out sensible values for both. Building the model is not your job. Feature work is automated. Bring a wider set of columns than you are sure about. Feature Refinery removes one column at a time, retrains, and measures what actually changed. You get back a reduced dataset and a per-column keep/remove decision you can show someone. It runs on CPUs. Training is a JVM process reading a CSV. No accelerators, no cluster, no queue for GPU capacity. Every model is deterministic. The same input produces the same decision every time, and you can replay any prediction. That matters when a decision has to be defended to a customer, an auditor or a regulator.How the pieces fit
1
Refine (optional but recommended)
Upload a labelled train file and a labelled test file. Feature Refinery returns a reduced pair
of files, the column decisions, and the performance target the refined data reached.
2
Build a dataset
Either hand the finished refinement straight over, or upload your own 80/20 split.
3
Train
The Crucible runs many attempts at once, scores each on the held-back 20%, and keeps the best
three as versions of one model.
4
Predict
Upload unlabelled rows, get a scored CSV back. Predictions always run against the version you
chose as champion.
Start here
Quickstart
From login to a first prediction. Around 15 minutes of your attention.
Input data
What your CSVs need to look like before you upload anything.
API reference
Base URL, auth, polling, error codes, and every endpoint.
Build an agent or CLI
A complete machine-readable brief. Point a coding agent at it and it can build a working client.
Access
Every customer gets their own tenant, isolated from every other. Your console and API addresses are issued when the tenant is created, and they follow this shape:
Examples throughout these docs write
<tenant>. Substitute the name you were given.
If your data cannot leave your own environment, MathFi.ai can run on your cloud account or private
cloud instead. That is arranged case by case rather than self-served, so talk to us at
[email protected] before planning around it.