The stroke patient who is missed, the heart-failure case that turns fatal, the radar signal that fails, the claim that should have been challenged — standard models miss them because they are rare. MathFi.ai is built for exactly those cases. Feature Refinery works out which of your data columns actually carry signal and drops the rest. Model Crucible takes the refined data, or your own CSV, runs multi-algorithm training and serves predictions behind an API. Same data in, same decision out, every time.
Under class imbalance, headline accuracy hides failure, so we measure the minority class: the stroke patient, the failing signal, the fraudulent claim, the case you actually care about. Here is how MathFi.ai performs against the five libraries most teams already run.
Of the patients who went on to have a stroke, found. XGBoost finds 30%, and F1 is a third higher than the next best method.
Of the heart-failure patients who die, flagged for escalation. The best other method flags 68.42%; XGBoost flags 52.63%.
Of rare warranty fraud surfaced for review, against 82 to 84% for the gradient-boosting methods. Those methods keep higher precision and F1 on this dataset.
| Decision | MathFi.ai · F1 % | Best other · F1 % |
|---|---|---|
| Stroke screening | 27.04 | 20.31 |
| Cardiac risk detection | 84.20 | 82.63 |
| Signal integrity / anomaly detection | 97.44 | 91.89 |
| Credit approval | 88.50 | 86.21 |
| Consumer default risk | 63.60 | 59.70 |
Highest minority-class F1 in 6 of 7 datasets, and the highest or joint-highest recall in all 7.
MathFi.ai benchmark, July 2026. Seven public datasets. Identical train and test splits. Five standard libraries as baselines.
Where a method beats us on a secondary metric, our study says so plainly.
See the full benchmark study, all seven datasets →Once a model is live, its data, its concepts and its performance all drift. For organisations that run hundreds of models, keeping them accurate often costs more than building them.
MathFi.ai is designed to need less of that work.
Cellular Balanced Learning (CBL), the architecture under MathFi.ai, has a small parameter space and only two hyperparameters. Drift has fewer paths into prediction errors, so a model is less likely to degrade abruptly when the data shifts.
Less recalibration per model means less MLOps effort. The saving grows with the number of models you run in production.
When drift appears, Feature Refinery analyses the new data and shows what changed in its structure and statistics. It then builds new representations, so you restore performance sooner than with a manual investigation.
Which claims are fraudulent. Which customers are about to leave. Which applicants will repay. Years of labelled history sit in your systems, and the patterns are in there.
It has meant specialists you cannot hire fast enough and a quarter of work per model: cleaning, feature engineering, tuning, validation, deployment, monitoring. Most organisations get a handful of models a year through that pipeline, if any.
The predictable fraud gets paid out. The savable customers leave. The good applicants get declined. Not because the signal is missing, but because the route from data to production has been too long to walk more than a few times a year. That is the bottleneck MathFi.ai removes.
MathFi.ai works on any decision that lives in labelled tabular data: risk, retention, operations, quality, safety. But the fastest payback is where the case you care about is rare and each miss is expensive. That is where standard models are weakest and MathFi.ai is strongest, so it is where most teams start.
Your source does not have to start as tabular data. Speech, images and sensor signals are all in scope once expressed as ML features: speech as time-series features, images as structured, labelled data capturing properties such as pixel intensity, texture and shape descriptors, and radar and sensor output as numeric signal descriptors.
Flag the rare, high-stakes cases that population-level screening misses.
Predict rare breakdowns from operational data before they become downtime.
Pay honest claims fast and route the small minority that deserve investigation.
Surface more fraudulent transactions, claims and applications for review, with fewer honest customers flagged along the way.
Approve more of the customers who will repay and catch more of those who will not, instead of trading one error for the other.
Identify the accounts likely to leave while there is still time to keep them, and aim retention spend where it changes the outcome.
The flow below is not a roadmap; it is a working day. You provide labelled data in the morning. Feature Refinery cleans it and selects the features that matter. Model Crucible takes the refined data, or your own CSV, and runs multi-algorithm training. MathFi.ai serves the predictions. By the end of the day your application, or your agent, is calling a live endpoint and getting deterministic decisions back.
A day is possible because of what is underneath. MathFi.ai runs on Cellular Balanced Learning (CBL), a genuinely novel machine-learning architecture our co-founders invented after decades of building production models the slow way. CBL is not a wrapper around open-source libraries; it changes how the model is built. Rather than searching across more than a hundred settings for a good decision boundary, CBL constructs dynamic cell boundaries that adapt to the structure of your data as it trains, and the relationships between them are learned inside the architecture.
That is why a CBL model exposes two hyperparameters where a gradient-boosted model exposes more than a hundred. The tuning decisions a specialist normally makes by hand, learning rate, depth, sampling, regularisation and the rest, are internalised in how CBL is built. The expertise sits in the architecture, not the operator, which is what lets a team without deep ML staff get a strong model, and lets a team that has one skip the tuning marathon.
It also earns its keep on quality. Most business prediction problems suffer from class imbalance: the cases that matter are a small minority of the data, and conventionally trained models sacrifice them for overall accuracy. Because CBL handles imbalance structurally rather than through oversampling patches, the same-day model is also the one with the highest minority-class F1 in six of our seven benchmarks. And it is a service, not a software project: MathFi.ai runs and serves the model for you, in our cloud or in yours, with every decision reproducible and auditable.
Read how CBL works in the docs →Most teams have tried at least one route to better predictions, and two things go wrong. The hard models, defeated by feature selection, more than a hundred hyperparameters and overfitting, often never reach production, so the value never lands. The ones that do ship are weakest exactly on the rare, costly cases. That gap, felt most by smaller teams frustrated at how little their current approach returns, is what MathFi.ai is built to close.
Doing this in-house means months of work per model and specialists who are hard to hire and harder to keep. Picking the right features, managing more than a hundred hyperparameters and fighting overfitting means many attempts never reach production at all, and the ones that do patch rare events rather than solve them, with oversampling tricks like SMOTE and class weights. MathFi.ai handles the rare cases structurally, in how the model itself is built.
The enterprise platforms are expensive and slow to adopt, and at the end of a long tuning run they hand back the same gradient-boosted models your team already runs, with the same blind spot on rare events. You pay platform prices to automate the status quo.
The new generation of tabular foundation models is genuinely interesting, but it is not deterministic. If you cannot reproduce a decision, you cannot audit it, explain it to a regulator, or debug it when it goes wrong. MathFi.ai gives the same decision on the same data, every time.
LLMs are probabilistic by design. That is the right property for language and the wrong one for a decision that has to be trusted: a lending call, a fraud flag, a churn intervention that must be right and repeatable. Use an LLM to talk. Use MathFi.ai to decide.
Model Crucible trains a model with each of its algorithms on the same data, so you can pick the balance between missed cases and false alarms that fits your decision.
Fewer missed cases and fewer false alarms at the same time; deterministic and auditable; your data team ships more models; and it is live the same day.
Finds which of your data columns carry signal, drops the rest and builds new representations from what is left.
Read the docs →Takes your refined data, or your own CSV, runs multi-algorithm training and serves predictions behind an API.
Read the docs →See every column of your data, its missing values and the balance of your target, before you train.
See more →You know the answers are in your data. Fraud losses, churn, bad debt: you can see the cost lines, but building models has meant hiring a team you cannot find or afford. MathFi.ai gives you a live prediction capability without the department. Bring labelled data; get a served model.
You have the team, and a queue of rare-event problems that resist the standard toolkit. MathFi.ai clears that queue: same-day models with minority-class recall your current stack struggles to reach, deterministic outputs your risk committee will accept, and your scientists freed for the work only they can do.
You operate where every decision may need to be explained, to a regulator, an auditor or a customer. Or you are building agents that act on predictions and cannot ship on a model that answers differently on the same input. MathFi.ai is deterministic end to end, so every decision is reproducible on demand.
MathFi.ai models are callable from your agents, copilots and applications through a documented REST API with an OpenAPI spec. Every prediction returns a deterministic, auditable decision your systems can log, replay and explain from the exact inputs the model saw. This is the difference between an agent you demo and an agent you deploy. Native MCP support is on the roadmap; today MathFi.ai is agent-usable through standard API integration.
Read the API docs →MathFi.ai was built by three operators who spent their careers shipping production machine learning and data systems at scale, and hit the exact wall that Cellular Balanced Learning was invented to remove. Between them, more than 20 granted patents.
Inventor of Cellular Balanced Learning. 25 years building machine learning, algorithms and patents at Fujitsu Laboratories, Tesco, import.io and Syncron. PhD from the University of Surrey.
LinkedIn17 years in backend and platform engineering, including at Tesco. Built the platform that takes CBL from research to production scale.
LinkedIn25 years leading product and technology at scale across Amazon, Tesco, GfK and startups.
LinkedInWe are choosing a small number of design partners. Bring a decision your current models get wrong too often, or a prediction you have never been able to build at all. We build a high-performance model on your own data and show you the lift against your own held-back validation data. This is for leaders who want to get past the tooling and straight to impact.
GenAI taught machines to talk. The next decade is teaching machines and agents to decide, and those decisions must be right, repeatable and auditable. MathFi.ai is the deterministic layer that GenAI calls when a decision must hold up: the prediction utility beneath the agentic enterprise.