Predictive AI Platform

Same day data to production model with unbeatable predictions

Structured data in, production predictions out. Same day. Two hyperparameters. Built to extend the people that already understand your data.

Register Your Interest
Independent Benchmark, May 2026

Outperforming State of the Art with expert-level accuracy and less tuning overhead.

MathFi.ai at 2 default parameters vs full grid search over 100+ parameters per competitor.

UCI Credit Approval: classifying consumer credit applications as approved or denied.

Method Accuracy F1 (approved) Precision Recall
MathFi.ai ★ 89.17% 88.50% 86.21% 90.91%
XGBoost 86.67% 86.21% 81.97% 90.91%
LightGBM 86.67% 85.96% 83.05% 89.09%
Random Forest 86.67% 85.96% 83.05% 89.09%
Logistic Reg. 85.00% 84.75% 79.37% 90.91%
CatBoost 84.17% 83.48% 80.00% 87.27%
Precision is the business-critical metric in credit scoring. A false approval (approving a bad credit risk) causes direct financial loss through default. MathFi.ai achieves higher precision than all five tuned competitors, approving fewer bad risks at the same recall rate as XGBoost.

★ MathFi.ai: default values (Buckets: 20). All other methods: full grid search. UCI Credit Approval dataset, 546/120 split, 120 test rows (65 denied / 55 approved). Independent benchmark, May 2026.

Same day
From raw data to first production prediction
2
Hyperparameters, vs 100+ for XGBoost and LightGBM
CPU
Commodity hardware. No GPU required, no cold starts
Deterministic
Same input, same prediction. Auditable by design
Who it's for

Built for teams that need results today.

Unlike conventional AutoML platforms that demand hundreds of configuration choices and long onboarding cycles, MathFi.ai is designed to fit into the teams, delivery motions, and systems you already run.

Mid-market data leaders

Bring predictive AI into the operating team

  • Your analysts and business teams already know the data, edge cases, and decisions that matter.
  • Load a CSV, set a target column, and produce a production-grade model the same day.
  • Use MathFi.ai to compress specialist engagement cycles without losing business ownership of the problem.
  • Your team iterates with clearer feedback loops, while experts stay focused on judgment, governance, and higher-value work.
Real case: 256 models across 16 countries delivered after a previous delivery path had stalled.
Analytics practices and enterprise DS teams

Scale model delivery without scaling tuning burden

  • When a prediction problem multiplies across business units or geographies, hyperparameter overhead becomes unmanageable.
  • 2 hyperparameters per model instead of 12: 256 models becomes 512 parameters to manage, not 3,072.
  • Data scientists stay on data quality, model governance, and novel problems instead of repeated commodity classification setup.
  • Deterministic design means drift response is simpler: fewer variables, more predictable behavior.
Real case: 3,072 hyperparameters reduced to 512 across a 16-country global deployment.
Regulated industries and agentic platforms

Predictions your systems can act on and your teams can verify

  • Regulated industries need decisions that are reproducible and explainable.
  • Agentic systems need a deterministic prediction layer alongside probabilistic language models.
  • MathFi.ai delivers mathematical output with a full audit trail.
  • Same input, same prediction, always. Not a configuration option: a property of the architecture.
Applied in financial services, industrial IoT, and clinical decision support.
Cellular Balanced Learning

A novel ML architecture, designed from the ground up.

Not a wrapper around open-source ML. Built for the cases where XGBoost, LightGBM, and AutoML create too much tuning, infrastructure, or delivery friction.

1

Traditional algorithms search. CBL constructs.

  • Traditional ML tools hunt for good decision boundaries by sampling combinations of 100+ parameters: learning rate, depth, regularization, subsampling, and more.
  • CBL doesn't hunt. It constructs dynamic cell boundaries across a flat cellular structure that adapt to your data's distribution as training progresses.
  • The adaptive structure optimises a powerful multi-dimensional predictive fitness function, reducing the amount of manual specification normally spread across 100+ parameters.
2

Rare events get their own cellular pattern.

  • Many standard workflows can be pulled toward the majority class. When 98% of rows are "normal", headline accuracy can hide weak minority-class performance.
  • Addressing this often means SMOTE, class weights, custom samplers, and manual tuning per dataset.
  • CBL's cell structure preserves rare-class examples instead of averaging over them.
  • Every class has the chance to get its own cellular pattern during training. Imbalance is handled structurally, not patched in after.
3

Mathematical output, not probabilistic.

  • Many high-performing ML workflows depend on random seeds or stochastic training choices.
  • For agentic systems and regulated industries, decisions are easier to govern when repeated runs are reproducible.
  • While MathFi.ai deploys random seeds, the seed dependency is insignificant for medium to large datasets: same input, same prediction, always.
  • This is not a configuration option; it is a property of the architecture. For smaller datasets, the prediction change pattern is cyclic and reproducible.
Architecture

Why 2 hyperparameters are enough.

MathFi.ai incorporates numerous internal parameters. As a platform, it optimizes these behind the scenes so teams can work with the 2 exposed hyperparameters that matter operationally.

  • 1
    Number of Buckets controls the granularity of the cellular structure. Applies to all data formats. Changing it has no impact on the other internal algorithm classes.
  • 2
    Scaling Factor applies to numeric and mixed numeric-categorical datasets. Has no effect on the bucket-based algorithm class.
  • ✓
    An internal selection layer automatically identifies which of the competing algorithmic pipelines performs best for your dataset. No manual algorithm selection needed.
  • ✓
    Training completes in minutes rather than hours. The linear architecture avoids the energy-intensive GPU compute required by deep learning and neural networks, running on commodity four-core CPUs.
  • ✓
    With only 2 hyperparameters, the tuning loop can be automated by an AI agent while data teams keep control of targets, validation, and deployment rules.
Advanced Performance Threshold (APT)

Stop training when you hit your target.

Set a performance threshold. MathFi.ai stops training when it's reached, avoiding unnecessary compute on models that already meet spec. For organizations running many models on weekly retraining cycles, this converts unpredictable cloud spend into predictable cost.

Real-world case

When 2 hyperparameters became a business requirement

  • 16 countries, 4 business divisions, 4 product lines: 256 predictive models needed.
  • Previous approach: up to 12 hyperparameters per model. Two months projected for initial delivery, before ongoing drift monitoring.
  • MathFi.ai: 2 hyperparameters per model, same scope.
3,072
hyperparameters with previous approach (256 x 12)
512
hyperparameters with MathFi.ai (256 x 2)
  • Data science team shifted more effort from repeated retuning to data quality and new business problems.
  • Post-deployment: drift in any market required reviewing 2 parameters per model, not 12.
Pair probabilistic agents with deterministic predictions.

Robust agentic systems combine probabilistic language models with deterministic systems of record. MathFi.ai is the deterministic prediction layer your agents can call inside governed workflows. Mathematical output. Full audit trail. Reproducible behavior by design.

Results

Strong fit for regulated and data-intensive industries.

Problem-agnostic. Benchmarked on real-world datasets across financial services, industrial operations, and healthcare.

Financial Services

Credit approval and financial risk

In credit scoring, precision matters more than recall. A false approval (approving a bad risk) causes direct financial loss through default. MathFi.ai leads all five benchmarked competitors on both F1 and precision, approving fewer bad risks at the same recall rate as the best competitor.

89.17% accuracy, 86.21% precision on held-out test set
Industrial IoT

Anomaly detection in operations

Gas plant and industrial sensor data presents extreme class imbalance: anomalies may represent under 6% of readings. CBL handles this structurally, without SMOTE or manual class weight configuration. Zero missed detections and zero false positives on the held-out test set.

100% F1, 100% precision on gas plant IoT benchmark
Healthcare

Clinical classification and early detection

Parkinson's telemonitoring presents imbalanced class distributions with rare minority healthy cases: 82% Parkinson's prevalence in the dataset. MathFi.ai correctly identified all healthy subjects with no false alarms in this benchmark, matching the top tuned competitors at default hyperparameters.

100% recall on minority class, Parkinson's benchmark
Team

Built by people who lived the problem.

The founding team created Cellular Balanced Learning and has built production ML and data systems at Amazon, Tesco, and global financial institutions. Over 20 granted patents between them.

SA
Saied Abedi
CEO & Co-Founder

ML researcher and algorithm architect. Inventor of Cellular Balanced Learning. PhD-level background in predictive AI and algorithmic design. Spent years building production ML at scale before creating CBL.

SO
Sean O'Neill
Chairperson / CPO & Co-Founder

Enterprise AI product and delivery leader. Decades of experience shipping production AI systems across retail, financial services, and global operations at scale.

DF
David Fernandez Morenza
CTO & Co-Founder

Large-scale data engineering and platform architecture. Built data infrastructure supporting services that drove billions in revenue.

Register Your Interest

Bring your use case.

We have a waiting list for Beta testing the MathFi.ai platform. Register your interest here, and we will contact you when a spot opens.