Platform · Model anatomy

Every layer, weight and bias, drawn and replayed.

The dashboard draws the structure you uploaded and walks a prediction through it in your browser, from the stored weights, with every step confirmed against the API's own trace. Below, a credit default forest you can walk one client at a time, then a network trace from the API reference, one operation at a time.

Model anatomy · credit-default-forest

Walk a real forest, one client at a time.

A random forest that scores card clients for default, drawn the way the dashboard draws it: ten inputs, one of its thirty trees, and the prediction. Pick a client the model never saw, or move an input, and the path moves. Every number is worked out in your browser from the model's own splits and leaves.

Model anatomy

Card default · random forest

Forest of 30 trees. The scene draws tree 1 of 30; a trace walks every tree and averages the leaves.

Family
Ensemble
Trees
30
Nodes
1,620
Fidelity
Exact
Input
Client

utilization 0.9388
from bill_amount ÷ credit_limit

Tree 1 of 30
How this prediction is madetree 1 of 30 · root to leaf

    Replayed in your browserfrom the forest's 1,620 nodes. For all 6,000 held-out clients this replay gives the same P(default) as scikit-learn, and for the four clients above the API's own trace takes the same path through every tree and returns the same probability. The data is Default of Credit Card Clients from the UCI repository (Yeh & Lien, 2009): 30,000 card clients of a bank in Taiwan, 24,000 of them used to train the forest. Moving an input makes a what-if, not a client in the data.

    Model anatomy · a real trace

    Follow one prediction through every weight.

    A small churn network from the Kaimei API reference scores one customer. Scroll to walk the prediction through it one operation at a time, with every number written out as it is computed.

    POST /v1/models/churn-net-docs/structure/trace

    1. 01 / 07The network

      One network, one customer.

      churn-net-docs is the example network in the API reference. All 27 of its parameters are uploaded, so every weight below is the one that computes the score. One customer waits at its inputs.

      Request body

      {"input": {"tenure_months": 3,
                 "monthly_spend": 91.5,
                 "support_tickets": 4}}
      • 11 nodes
      • 27 parameters
      • 4 layers
      • fidelity exact
    2. 02 / 07Standardise

      Standardise each input.

      The model stores a centre and a scale for every input. Three months sits 21 below the centre of 24; divided by the scale of 18, it becomes −1.167.

      tenure_monthstenure(3−24)/18=−1.1667 monthly_spendspend(91.5−62)/30=0.9833 support_ticketstickets(4−1.2)/1.6=1.75

      The trace's own line tenure months: (3 − 24) / 18 = -1.167.

    3. 03 / 07Hidden layer

      Four weighted sums.

      Each hidden unit weighs the three standardised inputs and adds a bias. The edges carry the products: blue raises a unit, pink lowers it, thicker moves it more.

      W
      0.8−0.30.5 −0.60.90.2 0.10.4−0.7 0.3−0.20.9
      ·
      x
      −1.16670.98331.75
      +
      b
      0.1−0.20.050
      =
      h
      −0.25331.735−0.89831.0283

      h₁ = 0.8·(−1.1667) − 0.3·0.9833 + 0.5·1.75 + 0.1h₁ = −0.9333 − 0.295 + 0.875 + 0.1 = −0.2533

      h₂ = −0.6·(−1.1667) + 0.9·0.9833 + 0.2·1.75 − 0.2h₂ = 0.7 + 0.885 + 0.35 − 0.2 = 1.735

      h₃ = 0.1·(−1.1667) + 0.4·0.9833 − 0.7·1.75 + 0.05h₃ = −0.1167 + 0.3933 − 1.225 + 0.05 = −0.8983

      h₄ = 0.3·(−1.1667) − 0.2·0.9833 + 0.9·1.75 + 0h₄ = −0.35 − 0.1967 + 1.575 + 0 = 1.0283

    4. 04 / 07ReLU

      Two units switch off.

      ReLU keeps positive values and turns negative ones into zero. Units 1 and 3 fall below zero for this customer, so their outgoing weights play no part in the score.

      relu(−0.2533) =0unit 1 off relu(1.735) =1.735 relu(−0.8983) =0unit 3 off relu(1.0283) =1.0283

      What the trace hands the output "input": [0.0, 1.735, 0.0, 1.0283]

    5. 05 / 07Output

      One logit, then a probability.

      The output unit weighs the two live units and adds its bias of −0.3, giving 2.3927. The sigmoid maps that to 0.916, above the 0.5 threshold, so the label is churn.

      z = −1.2·0 + 0.9·1.735 + 0.7·0 + 1.1·1.0283 − 0.3

      z = 1.5615 + 1.1312 − 0.3 = 2.3927

      p = σ(2.3927) = 0.91626639

      0.9163 ≥ 0.5 threshold churn

    6. 06 / 07Drivers

      What moved the score.

      The trace names the nodes that moved the final score, and by how much. The hidden layer pushed it up by 2.6927; the bias pulled it down by 0.3. By size, that is 90% and 10%.

      fromdeltashare hidden__relu+2.69270.8998 bias−0.30.1002

      Together 2.6927 − 0.3 = 2.3927 → σ → 0.91626639

    7. 07 / 07The API

      The same walk, in one call.

      The dashboard's anatomy view replays this walk in your browser, from the stored weights, with every step confirmed against the API's own trace.

      trace = client.models.trace_structure(
          "churn-net-docs",
          input={"tenure_months": 3,
                 "monthly_spend": 91.5,
                 "support_tickets": 4})
      print(trace.output)
      # {'label': 'churn', 'probability': 0.9163, …}

    Every number is from the churn-net-docs example in the Kaimei API reference (its spec, weights and trace response), for the reference's example customer. Values the reference elides with “…” are recomputed from the same weights, and every value it does print matches. Products are rounded to four decimals; sums use full precision. The drawing and its timing are made for this page; the dashboard draws the same structure in its own anatomy view.

    In the dashboard

    The same walk, for any model you upload.

    • Trees, forests, linear models and MLPsDrawn node by node, with the path a prediction takes highlighted.
    • Keras and PyTorch networksLayer graphs including convolutions, pooling, residual adds and concatenations.
    • Large models stay readableRepeated blocks fold into one, so encoder.layers.0 to 47 becomes a single ×48 block, and tensors load only when you open them.
    Kaimei model anatomy view of the card default random forest: ensemble, fidelity exact, with its ten inputs, the first of thirty trees and its labelled splits, and the prediction drawn as a 3D scene above a replay timeline

    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.