Solutions
Built for the models that carry consequences.
Credit decisions, trading signals, premiums, discharge plans, load forecasts. Wherever a model's output moves money or changes what happens to a person, someone has to answer for it. Kaimei gives that person the reasons.
Industries
Seven places a model has to answer for itself.
Each page works one real example end to end: the explanation of a single case, what the model learned, and the day production stopped looking like training.
Banking and lending
Underwriting, adverse action reasons, validation evidence and drift, for scorecards and machine-learning models alike.
- credit scoring
- adverse action
- model validation
30,000 card clients · exact reasons in 167 ms
See the worked exampleCapital markets
Signal attribution, regime drift and decay monitoring for alpha, risk and pricing models built on fixed features.
- signal attribution
- regime drift
- model validation
55.2% hit rate in sample, 50.3% after: decay, attributed
See the worked exampleInsurance
What each rating factor does, why a quote came out as it did, and when a new channel changes the book.
- pricing
- underwriting
- claims triage
678,013 motor policies · 9× lift across deciles
See the worked exampleHealthcare
Why a patient was flagged, what an imaging model looked at, and whether today's patients match the validated population.
- readmission risk
- imaging
- site validation
99,343 hospital encounters · a new ward flagged in one batch
See the worked exampleTechnology
One layer across churn, fraud, pricing and growth models, with versions, shadow scoring and alerts that name the input.
- churn
- fraud scores
- ML platforms
7,043 subscribers · one change takes 78% to 18%
See the worked exampleEnergy and industry
What each forecast leaned on, in megawatts, and an alert when the inputs leave the range the model learned in.
- load forecasting
- predictive maintenance
- visual inspection
8,568 hours of city load · error doubles when July drifts
See the worked exampleResearch and teaching
Exact, seeded and exportable results, published methods named for every number, and a model anatomy built for teaching.
- reproducibility
- teaching
- open client
6 worked examples on public data · Apache 2.0 client
See the worked exampleHow the examples were made
Real data, real models, real output.
- Public dataReal card clients, hospital encounters, motor policies, industry returns, city load and subscribers. Each dataset is cited and downloadable.
- An ordinary modelA random forest, gradient boosting or a neural network, trained the usual way and scored on rows it never saw.
- One uploadSent to Kaimei with the Python client, exactly as your team would send its own.
- Read backEvery attribution, counterfactual, insight and drift figure is what the API returned. No invented customers, no illustrative metrics.
By role
One record, read by everyone who answers for the model.
| If you are | You are asked | Kaimei gives you |
|---|---|---|
| Model risk and validation | Is the model conceptually sound, stable and still fit for use? | Importance, behaviour curves, learned rules, interactions and robustness, rerun with the same seed for the same result. |
| Compliance, audit and legal | Why was this decision made, and can you show it later? | A stored explanation per decision, retrievable by your own id, with the model version and reference population. |
| Data science and engineering | What changed, and what broke? | One upload call, version comparison, shadow models, and drift alerts that name the input. |
| Operations and the front line | What did the model see in this case? | A dashboard where any prediction can be opened, an input changed, and the attributions watched as they move. |
| Executives and boards | Which models are we running, and are any drifting? | A project view of every model, its traffic, its explained share and its open alerts. |
Deployment
Three ways to connect a model.
Chosen per model, by how much of it may leave your environment.
Upload the model
The client reads the fitted estimator in your process and sends its structure and weights. Kaimei runs it to explain it, and never stores the original file.
client.models.upload(model, …)Keep the model at home
Point Kaimei at a scoring endpoint you host, or send attributions you compute. The model never leaves your network.
inference="remote"Run Kaimei yourself
The API runs on your own infrastructure against your own database, for data that may not leave at all.
Self-hostingRules and guidance
Evidence for the reviews you already face.
| Area | Rules and guidance the pages map to |
|---|---|
| Model risk | US interagency model risk guidance of April 2026 (SR 26-2, which replaced SR 11-7); PRA SS1/23; OSFI Guideline E-23. |
| Consumer credit | Equal Credit Opportunity Act and Regulation B adverse action notices; GDPR Articles 15 and 22. |
| AI regulation | EU AI Act high-risk requirements for credit scoring and for life and health insurance pricing, applying from 2 December 2027. |
| Insurance | NAIC Model Bulletin on AI systems; Colorado Regulation 10-1-1; New York DFS Circular Letter No. 7 (2024); Solvency II internal models. |
| Markets | MiFID II RTS 6 on algorithmic trading; PRA SS5/18. |
| Healthcare | FDA guiding principles for machine learning in medical devices; EU AI Act requirements for medical devices from 2 August 2028. |
| Frameworks | NIST AI Risk Management Framework 1.0; ISO/IEC 42001. |
Kaimei produces evidence for these processes. It does not make a model, a notice or a filing compliant, and nothing on this page is legal advice. Rules are cited as they stood in November 2026; check what applies to you now.
Guides for risk and compliance teams: Reading an explanation, Adverse action reasons, Model risk management, The EU AI Act.
See it on your own model.
Tell us which model keeps you up at night. We will show you its first explanation.