Machine learning

Models that learn from your data to classify, score, detect and recommend, deployed and monitored in production.

Model card

Subscriber churn risk, v1.3

Example of the model card delivered with every model

1. Model details

Gradient-boosted trees (LightGBM), version 1.3, trained March 2026 by the data analytics team. Scores each subscriber's likelihood of cancelling in the next 60 days.

Start a project

What we build

Machine learning solutions

We build custom machine learning models on your data and put them to work in your products and processes: scores, predictions and recommendations delivered through an API, a scheduled job or the tools your teams already use.

Fraud detection
Score payments, claims or sign-ups for fraud risk as they happen, and send the riskiest to review.
Churn prediction
Flag the customers most likely to cancel, with the reasons, so retention teams can act first.
Recommendation engines
Product, content and next-best-action recommendations from behaviour and similarity.
Lead and account scoring
Rank leads and accounts by likelihood to convert or grow, inside your CRM.
Anomaly detection
Catch unusual transactions, sensor readings or system metrics as they happen.
Customer segmentation
Group customers by behaviour and value for targeting, pricing and planning.

An analyst studying image data across two monitors

How to get started

Three steps to a plan for machine learning

1Tell us what you needUse the project form or book a call. A few sentences about the goal is enough to start.
2Free technical consultationWe go through your goals, users, existing systems and constraints with you.
3Your planA detailed plan covering the right tech stack, architecture, timeline and budget. Then you decide.

Start a project Ask a question first

How it works

Five kinds of model we build

Each answers one business question and returns an output your team or systems act on. Choose one to see an example.

Supervised

“Will this transaction turn out to be fraud?”

Model output

Card not present, new device

Amount: $1,240
Time: 03:12
Fraud probability: 0.91

Hold for review

It outputs: A category, or the probability of each one.
Measured by: Precision, recall and ROC AUC, with the threshold set by the cost of each kind of mistake.

An explanation with every prediction

Every score we deliver comes with the factors behind it. Choose a customer, then a factor, to see what the retention team sees and can do.

Predicted churn risk: 39%

Raises risk Lowers riskPercentage points of churn risk. Choose a factor.

Factor

Month-to-month contract

Raises risk by 14 points

What it means

No term commitment, so the customer can leave at any time without a fee.

Retention action

Offer an annual plan at a lower monthly price.

Our approach

How the work runs

Discovery
Agree the decision the model supports, the success metric and the current process it has to beat.
Data assessment
Check the history, quality and access of your data, then build the features the model learns from.
Model development
Train, tune and compare candidate models, then test the best on data held back from training.
Deployment and integration
Serve the model as an API or scheduled job and connect it to the systems where predictions are used.
Monitoring and support
Track accuracy and drift, retrain on a schedule or when alerts fire, and support the model after launch.

What we'll need from you

Having these ready keeps the work moving.

Data access
Read access to the systems that hold the history and outcomes, or extracts to agreed specifications.
Domain expertise
People who know the process to explain the data, sanity-check features and review results.
A decision owner
The person who will act on the predictions and agrees the success metric.
A route to production
Contact with the teams that own the systems where predictions will appear.

Who's on the project

Our team, working with decision owner and domain experts from yours.

¹, ¹, ¹, ², ²

¹ Mayrian ² Your organization

Data scientist: Frames the problem, builds features and models, and explains results.

Services

Machine learning services

Machine learning consulting
Use-case selection, a data assessment and a feasibility check before you commit to a build.
Custom model development
Features, training and evaluation against your current process, documented in a model card.
Deployment and integration
Models served as a real-time API or scheduled batch job, connected to your product, CRM or data warehouse.
MLOps and model support
Automated training pipelines, a model registry and CI/CD for models, with monitoring, retraining and support after launch.

Deliverables

What you receive

Appendix A. What you receive

All code, data and documentation are handed over in your accounts and repositories.

In your hands

What the documentation looks like

Every project ends with documents your team can run with. Here is an excerpt of one of them.

Model card

Payment fraud classifier, v2.3

Gradient-boosted trees · retrained monthly

Intended use

Rank card-not-present payments for review. Not for declining payments without review.

Training data

18 months of payments with chargeback labels; last 3 months held out.

Evaluation on held-out data

Measure	Model	Current rules
Recall at 1% review rate	71%	48%
Precision at 1% review rate	38%	21%
ROC AUC	0.94	—

Limitations

Lower recall on new merchant categories; monitor drift monthly with PSI.

Measuring success

How success is measured

What we report on in machine learning projects. Which measures apply, and their targets, are agreed with you at the start.

Table 2. What we report, and when. Targets are agreed with you at the start.

Measure	Reported
Precision and recall	Held-out test, then monthly
ROC AUC	Held-out test
Calibration	Held-out test
Lift over baseline	Held-out test, then quarterly
Performance across groups	Held-out test, by group
Drift	Continuously in production

Measure

Precision and recall

How many flagged cases were right, and how many real cases were caught, at the threshold agreed with you.

Reported

Held-out test, then monthly

Tuned to your costs

We set each model's decision threshold from what a missed case and a review cost you. Drag it, or enter your own costs.

Decision threshold0.35

Cost of a missed fraud$250

Cost of a review$6

FlaggedNot flaggedFraudGenuine

False negative: Fraud that gets through, usually ending in a chargeback and fees.

Precision

Recall

100%

Daily cost

$11,553

Lowest daily cost at 0.50: $2,396

How to start

From first call to production

Start where you are. Each step ends with a decision, so you commit to the next one only when it makes sense.

Protocol

How an engagement runs

Each step ends with a decision on whether to continue

Free technical consultation

One or two sessions

Talk through the goal, the data you have and the systems involved.

Outputs

(a) A shortlist of use cases, ranked by value and feasibility
(b) A recommended next step
(c) A plan covering stack, architecture, timeline and budget

Book the consultation

Estimate the value

What it could be worth to you

Enter your own figures. The formula is shown, and the estimate can go with your enquiry.

Estimate

Fraud losses avoided

Machine learning · from your own figures

Losses avoided = Annual fraud losses × share caught earlier(1)

Fraud losses a year$1,200,000Share of today's losses caught earlier25%

Result

$300,000

a year

Add to my enquiry

Build or buy

When you don't need
a custom build

Part of the free technical consultation: when an existing product covers the need, we recommend it instead of a custom build. These are the options we weigh, alongside the tools you already have.

Related work

Existing products that may be enough

[1]Scoring built into your CRM or payments platform, such as HubSpot predictive lead scoring or Stripe RadarWhen the question fits what the product already scores, using data held in that product.
[2]AutoML tools such as Azure Automated ML, Amazon SageMaker Canvas or AutoML on Google CloudWhen a standard tabular problem, with a clean dataset and an analyst to run it.

Our approach

When a custom build is worth it

The data spans several systems of your own
Decisions need explanations, fairness checks or model-risk documentation
Predictions must run inside your product or processes
The problem is specific enough that generic scores miss it

Governance and monitoring

Controls built into every model

What we set up on every machine learning project, so each model is documented, reviewable and monitored.

1Before launch
- Evaluation reportAccuracy against your current process on held-out data, with results by customer group to catch bias.
- ExplanationsThe factors behind each prediction, shown to reviewers and summarized in the model card.
- Model card and sign-offIntended use, data, limits and results documented, and approved by your decision owner.
2At launch
- Model registryEvery model version recorded with its training data and results, with rollback to an earlier version.
- Access controlRole-based access to models, features and predictions, in your own cloud account.
- Prediction loggingEach prediction stored with its inputs and model version, for review and audit.
3After launch
- Drift and accuracy monitoringDashboards and alerts when incoming data or accuracy moves away from what the model was trained on.
- RetrainingNew versions trained on a schedule or when alerts fire, each evaluated before it replaces the last.
- Human reviewQueues for people to review borderline or high-impact cases before action is taken.

In your industry

Where it applies

What this work delivers in the industries we serve.

Technologies and standards

Chosen for your project

Built on the cloud you already use. Choose yours to see the services involved; we recommend the full stack in the free technical consultation.

Table 2. Managed services for each layer, by cloud. The highlighted column is the one you chose.

Layer	AWS	Azure	Google Cloud
Training and experiments	Amazon SageMaker AI	Azure Machine Learning	Gemini Enterprise Agent Platform: training
Feature store	SageMaker Feature Store	Azure ML managed feature store	Gemini Enterprise Agent Platform: feature store
Pipelines	SageMaker Pipelines	Azure ML pipelines	Gemini Enterprise Agent Platform: pipelines
Serving	SageMaker endpoints	Azure ML online endpoints	Gemini Enterprise Agent Platform: endpoints
Monitoring	SageMaker Model Monitor	Azure ML model monitoring	Gemini Enterprise Agent Platform: model monitoring

Also runs on any of the three: Databricks, MLflow, Snowflake ML, scikit-learn, XGBoost, PyTorch.

Frameworks

scikit-learn
XGBoost
LightGBM
PyTorch
TensorFlow

MLOps

MLflow
Amazon SageMaker AI
Gemini Enterprise Agent Platform
Azure Machine Learning
Kubeflow

Explainability and monitoring

SHAP
Evidently AI
MLflow Model Registry
Model cards

Data and features

Databricks
Snowflake
dbt
Feast

Next section

Computer vision

Models that detect, count and inspect, tested in real conditions.

Previous: Predictive analytics

Data Analytics & AI

5 Computer vision

What's in the image?

Questions

Common questions
about machine learning

How much data do we need?

We confirm it for your use case before any build starts. The data assessment checks how many past examples of the outcome you have, how far back they go and how complete they are. If there isn't enough yet, we tell you what to start collecting.

Can you explain why a model made a decision?

Yes. Every model we deliver comes with the factors behind each prediction, from an interpretable model or SHAP values. Reviewers see them beside the score, and the model card summarizes them.

Real-time or batch predictions?

Both. We deliver real-time predictions through an API for decisions made in the moment, such as fraud checks at checkout, and scheduled batch scoring when predictions can be prepared in advance, such as a weekly churn list.

Do you work with the platforms we already use?

Yes. We build on AWS, Azure or Google Cloud, and on Databricks or Snowflake, in your own accounts, so the models, data and pipelines stay yours.

Do you support models after launch?

Yes. We monitor accuracy and drift, retrain models on a schedule or when alerts fire, and keep supporting them, or hand everything over to your team with the documentation to run it.

How is a project priced?

Well-defined scopes are delivered as fixed-price engagements; when requirements are still evolving, we provide a dedicated team instead. Either way, the free technical consultation ends with a plan covering tech stack, architecture, timeline and budget, so you know the cost before work starts.

Machine learning

Machine learning solutions

Fraud detection

Churn prediction

Recommendation engines

Lead and account scoring

Anomaly detection

Customer segmentation

Three steps to a plan for machine learning

Five kinds of model we build

An explanation with every prediction

How the work runs

Discovery

Data assessment

Model development

Deployment and integration

Monitoring and support

What we'll need from you

Who's on the project

Machine learning services

Machine learning consulting

Custom model development