Supported models and what's coming
A plain list of the models Kaimei explains today, the ones it explains through your own scoring endpoint, and the models and features coming soon.
#Explained today
Kaimei explains models that make one prediction from a fixed set of inputs. Each input is a number, a category, a yes/no flag, or an image.
| Kind of model | Examples | How it connects |
|---|---|---|
| Yes/no and multi-class decisions on structured data | Credit approval, fraud flags, churn, claims triage | Upload it from scikit-learn, Keras or PyTorch, or connect a scoring endpoint. |
| Numeric predictions on structured data | Demand forecasts from features, pricing, loss estimates | The same. |
| Image classification | Defect detection, document type from a scan, species | Upload a convolutional network, or connect a scoring endpoint. |
| Models Kaimei runs through your endpoint, on the same kinds of input | Non-linear SVMs, nearest neighbours, naive Bayes, Gaussian processes, tree models with categorical splits, models written in other languages | A scoring endpoint you host, or attributions you compute. Running more of these directly is coming soon. |
#Partly supported
- Networks with attention, recurrent or upsampling layers. Kaimei draws the architecture, and running those layers itself is coming soon. Until then, connect a scoring endpoint to explain the model, provided its inputs are structured data or images.
- Custom models. Kaimei runs tree ensembles of any size, linear models, neural networks and combinations of those itself; anything else, such as your own Python classes, needs a scoring endpoint.
- Very wide inputs. Above 96 inputs, the Shapley methods first find the inputs that matter and report the rest as one remainder. LIME and KernelSHAP stop at 4,096 inputs.
- Grad-CAM needs an image model whose layers were uploaded. Occlusion and LIME work for any image model.
#Coming soon: models and inputs
Kaimei explains one prediction from a fixed set of inputs today. These are the models and inputs we are adding, and what to use until each one ships.
| Coming soon | What Kaimei will explain | Until then |
|---|---|---|
| Text and language models | Sentiment, topic and document classifiers on raw text, with each prediction attributed to words and passages; large language models such as chat assistants, summarisers and code generators, through the score of a chosen answer. | Explain the structured features you derive from the text, such as scores or counts. |
| Generative models | Models that produce text, images or audio, through the score or likelihood of a chosen output. | Register a score you derive from the output, such as a quality or safety score, as a numeric model behind a scoring endpoint. |
| Raw sequences and time series | Models that read a history step by step, such as recurrent networks and temporal transformers, attributed per time step and per input. | A forecasting model that takes fixed features, such as lagged values and rolling averages, works today. |
| Several outputs at once | Multi-output regression, multi-horizon forecasts and multi-label classifiers, with every output explained in one call. | Register one model per output. |
| Boxes, masks and rankings | Object detection and image segmentation, explained per box or mask; recommendation and search ranking, explained per item's position. | Image classification works today, and a ranker's score for one item can be registered as a numeric prediction. |
| Audio and video | Attributions per time span, frequency band or frame. | A classifier on features you extract, such as spectral statistics, works today. |
| Clustering and reinforcement learning | Why a row falls in its cluster, and which observations drive a policy's action. | A model that outputs a single anomaly score can be registered as a numeric prediction. |
| ONNX and Hugging Face uploads | Models exported to ONNX, including models written in other languages, and Hugging Face models, uploaded and run by Kaimei itself. | Connect a scoring endpoint; the framework is recorded as a label. |
| More estimators, run directly | Tree models with categorical splits, scikit-learn pipelines with encoders, non-linear SVMs, nearest neighbours and naive Bayes, read and run by Kaimei. | Connect a scoring endpoint, or send attributions you compute. |
| Attention, recurrent and upsampling layers | Networks with these layers run by Kaimei, not only drawn. | Kaimei draws the architecture; connect a scoring endpoint to explain the model. |
#Coming soon across the platform
Beyond new kinds of model, these are the features we are building for the teams that explain, monitor and govern models.
| Coming soon | What it will do | Until then |
|---|---|---|
| Fairness and disparate-impact testing | Outcome rates, adverse impact ratios and calibration compared across the groups you define, with a check for inputs that act as proxies for a protected characteristic. | Explanations can help investigate a finding, but they are not a test. |
| Integrated Gradients and DeepLIFT | Gradient-based attributions computed by Kaimei for uploaded networks. | Store the attributions you compute beside the built-in methods. |
| Adversarial robustness | Gradient attacks such as FGSM and PGD, reporting how small a change flips each decision. | The robustness analysis measures decision flips under random noise. |
| Challenger benchmarks | A simple, interpretable model fitted to the same reference data, so validators can see what the complex model adds. | Register a challenger yourself and compare the two models' insights side by side. |
| Scheduled checks | Drift and performance analyses run by Kaimei on the schedule you set. | Run them from your own scheduler, one API call each. |
| Alert routing | Medium and high findings sent to email, Slack, Microsoft Teams, PagerDuty or a webhook. | Alerts appear in the dashboard and in the organization and project overviews. |
| Data quality checks | Missing values, out-of-range values, unseen categories and unit changes caught as predictions are logged. | Drift compares live inputs with the reference data, feature by feature. |
| Attribution drift | An alert when what the model relies on shifts, not only its inputs. | Compare the global importance from analyses run at different times. |
| Retention schedules | Records deleted automatically once they pass the period you set. | Records are kept until you delete them. |
| Policy rules beside the model | Cut-offs, overlays and manual decisions recorded with each explanation, so it reflects the whole decision. | Keep them in your decision system, under the same prediction_id. |
| Reason tables | Approved adverse-action wording kept in Kaimei under change control, so each explanation returns draft reasons and the inputs held for review. | Keep the table in your code, as in Adverse action reasons. |
| Documentation reports | Model documentation and validation reports drawn from the structure, analyses and insights, exported as PDF or Word. | Structure and insights export as JSON. |
| Single sign-on and provisioning | SAML and OpenID Connect sign-in through your identity provider, with SCIM to add and remove people. | Sign in with email and password, Google or GitHub. |
| Audit log | Who signed in, created a key, changed a model or deleted a record, and when. | Every prediction, explanation and analysis is stored with its time and model version. |
| ML platform integrations | Register models straight from MLflow, Amazon SageMaker, Vertex AI and Databricks. | Upload from your notebook or pipeline with the Python client. |
| An R client | The same calls from R, for statisticians, actuaries and teaching. | The REST API works from any language. |
#By design
- Cause and effect. An explanation describes how the model responds to its inputs, not what would happen in the world if an input changed.
- Serving your model. Kaimei calls your model, or runs an uploaded copy, to explain and analyse it; your own systems keep serving customers.
If your model, or a feature you need, is not listed here, tell us about it. It helps us decide what comes first.