mahdev

Risk and machine learning models that hold up in production.

mahdev is a data science consultancy. We design, build and validate pricing, risk and machine learning models, from raw data to a documented model your team can run, explain and defend.

Talk to us Our services
Illustrative smooth effect from a generalized additive model A curve showing how claim frequency changes with driver age: high for the youngest drivers, falling to its lowest point around age 50, then rising slightly for older drivers. A shaded confidence band is widest at the youngest and oldest ages, where there is less data.
Illustrative example. How claim frequency changes with driver age in a generalized additive model, with everything else held fixed. The band is the 95% interval; the marks along the axis are the data behind it. We build models where every effect can be shown, checked and explained like this.

Services

We work with insurance and financial services teams that need models they can trust with real decisions. We can deliver a model end to end or work alongside your team on the parts that slow it down.

Pricing and risk modeling

Frequency, severity and loss-cost models built with GLM, GAM and EBM. Rating factors, interactions and credibility-based groupings that hold up under review.

Machine learning models

Explainable machine learning for risk scoring, segmentation and cross-sell propensity, built so every prediction can be traced back to its drivers.

Data engineering

Scalable pipelines in PySpark and SQL, with automated data-quality checks that catch missing values, drift and broken joins before anything is modeled.

Model validation

Independent reviews of existing models: back-testing, stability checks, lift and Gini analysis, and clear findings your team can act on.

AI-assisted modeling

Workflows built on the Claude API that speed up coding, variable testing and model documentation inside your team, with a modeler reviewing every step.

Our toolkit: R and mgcv, Python and InterpretML, PySpark, SQL, KNIME and the Claude API.

How an engagement works

  1. Understand the decision

    We start with the decision the model will support and the data you already have, and agree on what success looks like.

  2. Build in the open

    Data pipeline, variable selection and model development, reviewed with your team at each step rather than revealed at the end.

  3. Validate

    Back-testing on held-out data, stability checks over time, and a plot for every effect so nothing in the model is a surprise.

  4. Hand over

    Documented code and models that your team can run, update and explain without us.

What you can expect

Models you can explain

We favor transparent models. When a more complex model earns its place, we show exactly what drives it.

Built for production

Code that runs on your infrastructure and your data volumes, not a notebook that works once.

Your data stays with you

We work inside your environment. AI tools only see code, aggregated outputs and documentation, never customer-level data.

Let's talk.

Tell us about the decision you want to model and the data behind it. We'll reply by email to set up a call.

Prefer email? destek@mahdev.org