Technology: The Engine

The infrastructure for
AI-native insurance.

Soldek.ai isn't an AI tool. It's the core risk engine and infrastructure layer insurers and AI-native insurers are built on.

Architecture

Input. Logic. Output. One engine.

01 · Input

Submission

data · context · niche
02 · Core

Soldek.ai Risk Engine

frequency-severity GLM · real data + expert priors · XAI
03 · Output

Priced quote

bound · explained · audit-ready
Foundations

Four properties. Non-negotiable.

Compliance, transparency, ownership, openness: built into the core, not bolted on at the end.

01 · XAI

Explainable by design

Every decision the engine makes is traceable. Underwriters, auditors and regulators see why the model priced what it priced, not a black box, not a guess.

Decision trace Audit-ready
02 · EU AI Act

Designed for the EU AI Act

Built from day one with the EU AI Act high-risk regime in mind: model cards, bias and proxy-discrimination scans, and fundamental-rights impact assessments ship with the engine. Compliance work isn't an upgrade. It's the foundation.

EU AI Act GDPR Model cards
03 · Own IP

Our engine. Our IP.

The risk engine, pricing models and policy stack are proprietary: a frequency-severity GLM, fitted on real market claims data where it exists (motor: 5,100,000+ public policies from six markets; crop: 5,900,000+ US policy-years of storm-, drought- and hurricane-driven experience; flight disruption: 7,000,000+ US flight-legs behind a parametric delay cover; motorcycle, health and commercial property: real public books) and cold-started on expert priors from 30 years of senior underwriting elsewhere, recalibrating on real claims experience. Each customer routes to the most specific fitted model: country, region, world. The conversational layer uses OpenAI, pricing never comes from the LLM.

Owned IP Real data + priors
04 · Open infrastructure

Partners plug in. Data stays yours.

The engine is built as infrastructure. Insurers and partners connect their claims experience: the engine is built to recalibrate on it, and the IP stays with the partner.

Partner-ready Data sovereignty
What it's trained on

Real market experience. Line by line.

Every bar below is real, public market experience the engine is fitted and backtested on — 18,000,000+ policies, policy-years and flight-legs from nine markets.

Flight disruption 7,000,000+ flight-legs
Every scheduled US flight of 2024 (US DOT/BTS On-Time) — parametric delay and cancellation cover
Crop 5,900,000+ policy-years
US federal crop book (USDA RMA) — storm, drought, hail and hurricane-driven losses
Motor 5,100,000+ policies
Six markets — France (freMTPL2), Belgium (beMTPL97), Italy (euMTPL), Brazil (brvehins1), Australia (ausprivauto), Norway (norauto) — plus a pooled world model
Health 176,000+ member records
European collective health book (euhealthinsurance) with per-member claims experience
Motorcycle 62,000+ policies
Swedish motorcycle book (swmotorcycle) with rider, vehicle and region factors
Commercial property 5,600+ policy-years
US commercial property fund (Wisconsin LGPIF) — buildings, contents and weather perils
Footballer TTD and PTD covers are priced deductively from the published UEFA injury epidemiology (incidence, lay-off distribution, career-ending hazard) and the contract terms; the other lines start from expert priors calibrated to published market benchmarks. All of them recalibrate on real claims as they flow in:
Footballer TTD & PTD Travel Contents Liability Cargo Business liability Accident & sports injuries Cyber
Validation

Out-of-sample backtests. Not projections.

Measured on the real books above, on data the models never saw during fitting.

7.5× → 1.8×
Mispricing spread across price deciles on the French motor book — 76% less cross-subsidy than flat pricing
26.6× → 1.4×
The same spread on the US crop book — 95% of the flat-pricing mispricing removed
0.93
Gini risk ranking on the crop book; 0.89 on the pooled world motor book, 0.79 on flight disruption and US commercial property
±1%
Actual-vs-expected accuracy on the two largest books (crop and world motor), out of sample

Every price the engine produces decomposes into named risk drivers — explainable AI by design, documented for the EU AI Act, with the full validation reports shipping alongside the platform. Your own claims data recalibrates the models; it stays yours.

The difference

General models guess.
Our engine calculates.

GLM
Frequency-severity pricing core: actuarial math, not an LLM
0
Pricing decisions taken by an LLM: the models set every price
30y
Senior underwriting expertise encoded as the engine's priors
XAI
Per-feature explanations · model cards · auditable decisions

Plug your data into the core.

Private demos for insurers, MGAs and partners. See how the engine prices your niche, and how it integrates.

Book a demo