Platform · Security

Private by design.

Kaimei never stores your original model file. Upload shapes only and the weights stay in your environment. Point it at your own endpoint and the model never leaves at all.

Private by design

What leaves your network, and what never does.

Kaimei never stores your training set or your model file. It keeps what you choose to send, model by model: the weights, only their shapes, or none of the model at all. Below, one model is connected each way: a hospital’s 30-day readmission score.

  1. 01Upload

    The default upload sends the structure, never the file.

    client.models.upload reads the fitted estimator in your process and extracts its layers and learned weights into a kaimei.structure.v1 spec. The original pickle, .h5 or .pt is never sent. Pass X and y and the reference rows you choose go with it, up to 5,000.

    For the hospital’s gradient boosting that was one 79 kB request: 150 trees, with every split and leaf value. Each reference row is the ten model inputs and the outcome, with no name, record number or date. Kaimei executes this model family, so explanations, predictions and analyses then run in the API.

    client.models.upload(
        model, name="30-day readmission",
        X=reference, y=outcomes,
    )
  2. 02Shapes only

    Keep the weights at home.

    With weights="summary" the client sends shapes and statistics instead of the weight values. The dashboard still draws the model from them. The weights never leave your environment.

    A tree model’s weights are its split thresholds and leaf values. Without them the hospital’s request fell from 79 kB to 4 kB: how many trees, how large each is, and one importance per input. Shapes cannot be executed, so explanations need a scoring endpoint. Logging predictions and storing your own attributions work without one.

    client.models.upload(
        model, name="30-day readmission",
        weights="summary",
    )
  3. 03Scoring endpoint

    Or keep the whole model at home.

    Register an HTTPS scoring URL. When Kaimei needs a prediction it POSTs batches of rows to your endpoint and reads the predictions back, so explanations, analyses and predict all work while the model stays on your infrastructure.

    Headers you set are stored and never returned. The endpoint has to be reachable from Kaimei’s servers. Explaining one discharge through the hospital’s endpoint took 8 requests carrying 1,031 rows. That estimate is sampled; run in Kaimei, the same explanation is exact.

    inference={"type": "remote",
               "url": "https://…/predict"}
  4. 04Your attributions

    Or send only the numbers.

    Already computing TreeSHAP or integrated gradients? Send the values to POST /v1/explanations. Kaimei does not query the model: it ranks them, records the method as provided and draws them with the same bars as its own.

    Shown: the readmission model’s own TreeSHAP values for one discharge, in log-odds, computed inside the hospital and sent in a request under 1 kB. Counterfactuals and analyses need a model Kaimei can run.

    client.explanations.create(
        model="readmission-30d",
        input=discharge,
        attributions=treeshap_values,
    )
  5. 05Monitoring

    Logging works in every mode.

    Your serving stack logs each prediction with its input, its output and your own id, up to 1,000 per call. Logging never needs Kaimei to run the model. A drift check does: it compares the latest 500 logged inputs with the reference rows, feature by feature, and scores both, on the uploaded model or through your endpoint.

    The hospital logged the same discharges in all four modes. After a batch of patients aged 85 and over, the check through its endpoint scored 1,000 rows and reported age PSI 6.73, significant.

    client.predictions.create_batch(
        model="readmission-30d",
        items=served,   # up to 1,000
    )
  6. 06Summary

    What crosses, mode by mode.

    Kaimei stores the feature schema you register, the reference rows you choose to upload, the layers and weights you choose to extract, and the predictions and explanations it produces. Every call is HTTPS to api.kaimei.ai/v1 with a bearer token, and API keys are stored hashed.

    For the hospital that is rows of ten model inputs with their outcomes or predictions, and as much of the model as it chose to send. Enterprise plans include a private deployment.

    What crosses your network boundary in each mode
    Leaves your network Upload Shapes only Endpoint Your attributions
    Model filenevernevernevernever
    Training setnevernevernevernever
    Weight valuessentstaysstaysstays
    Shapes and statisticssentsentoptionaloptional
    Feature schemasentsentsentsent
    Reference rowsoptionaloptionaloptionaloptional
    Logged predictionsoptionaloptionaloptionaloptional
    Kaimei runs ityesnovia your endpointno

    Reference rows are the rows you choose to upload, at most 5,000; here, a sample of the training encounters. An upload runs in Kaimei when its model family is one Kaimei executes. Logging never needs Kaimei to run the model; drift checks and other analyses do.

Schematic, not a network map. The model is the readmission example from the Healthcare page: gradient boosting on Diabetes 130-US Hospitals, 1999–2008 (UCI), a public, de-identified dataset. It was registered once per mode on a local Kaimei API, with the scoring service on the same machine. The request sizes, the 1,031 scored rows, the TreeSHAP values (prior_inpatient +1.77, prior_emergency +0.34, discharge +0.21 log-odds, from shap) and the drift figure are from that run. The traffic is staged: held-out discharges replayed as logged predictions, then a batch of patients aged 85 and over, as on the Healthcare page. The packets and their timing are illustrative.

In practice

The details a security review asks for.

Bearer tokensRequests and responses are JSON over HTTPS, authenticated with a bearer token. Authentication
Keys stored hashedAPI keys are stored hashed and cannot be recovered; the secret is shown once. API keys
Endpoint headersHeaders your scoring endpoint needs are stored and never returned. Scoring endpoint
Private deploymentEnterprise plans include a private deployment. Self-hosting

Coming soon

Coming soonSingle sign-onSAML and OpenID Connect sign-in through your identity provider. Email and password, Google and GitHub work today.
Coming soonUser provisioningSCIM, so people are added to and removed from Kaimei with your directory.
Coming soonAudit logWho signed in, created a key, changed a model or deleted a record, and when.
Coming soonRetention schedulesRecords deleted automatically once they pass the period you set. Until then, they are kept until you delete them.

Upload a model this afternoon.

The free plan covers 10 models, 500 explanations and 20 analyses a month. Your own model is one call away.