Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.mathfi.ai/llms.txt

Use this file to discover all available pages before exploring further.

Getting your first prediction typically takes under an hour once you have labelled CSV data ready. Training runs in minutes to a few hours depending on dataset size and complexity. Integration into existing workflows via the REST API can be completed in a day by a developer familiar with HTTP.
MathFi.ai is designed for classification problems on tabular data — both binary (e.g. fraud / not fraud) and multi-class (e.g. categorising outcomes into three or more groups). Regression problems can also be solved by converting continuous targets into discrete classes, as described in the Training guide. Real-world examples are available in the Use Cases section.
Pricing is usage-based and tailored to your organisation’s needs. Request access or contact us at [email protected] to discuss pricing for your use case.
The web dashboard requires no coding — you upload a CSV, configure a few settings, and download predictions. To integrate via the REST API, basic familiarity with HTTP requests (e.g. curl or any HTTP client) is sufficient. A Python SDK is also available for teams that prefer a programmatic interface.
MathFi.ai uses CSV as its input and output format, which is compatible with virtually every data tool and database. The REST API allows programmatic integration into existing pipelines, and the Python SDK makes it straightforward to embed MathFi.ai into Python-based workflows.
All datasets and trained models are stored securely in the cloud under your account. Access is protected by bearer token authentication with a 1-hour expiry. Contact [email protected] for questions about data retention, compliance, or enterprise security requirements.
When you upload a labelled CSV, MathFi.ai automatically splits it: 90% is used for training and 10% is held back as an unseen test set. Training performance measures accuracy on the 90% the model learned from; test performance measures accuracy on the 10% it has never seen. A healthy model has training and test performance that are close together — a large gap signals overfitting. See the Glossary for detailed metric definitions.
MathFi.ai evaluates model performance using a balanced metric that averages accuracy across each class, rather than counting overall correct predictions. This means a model must perform well on every class — not just the dominant one — which is critical for imbalanced datasets such as fraud detection or rare disease diagnosis. The platform automatically generates a balanced test set from your labelled data to ensure fair evaluation. See the Glossary for the full formula.
Yes. The MathFi.ai dashboard is designed for non-technical users: upload a labelled CSV, configure three plain-language settings, and the platform handles the rest. No knowledge of machine learning algorithms or statistics is required to run predictions. Data scientists can go deeper via hyperparameter tuning and the REST API when needed.
MathFi.ai exposes exactly three hyperparameters:
HyperparameterWhat it controlsDefault
Number of BucketsHow the dataset is divided into bins for training. A higher value can improve accuracy but may cause NaN predictions if set too high for small datasets.20 (range: 4–100)
Scaling FactorThe number of search cells explored during training — higher values search more broadly.19 (range: 8–499)
Performance ThresholdThe minimum accuracy a model must achieve to be accepted as a champion.0.8 (range: 0–1)
See the Hyperparameter tuning guide for a step-by-step approach to optimising these values.