MLDoo · Build & Explain
Status: PreviewBuild ML you can understand.
Develop, evaluate, explain, deploy and monitor machine learning models with interpretability at the center.
- What it is
- MLDoo is a machine learning studio. Design a model as a visual flow, train it on versioned data, read its metrics and explanations, register it, and deploy it behind an endpoint you can monitor.
- Who it is for
- Data scientists and ML engineers who have to justify a model to the people accountable for its decisions — and analysts taking their first step into modelling.
- Status
- Preview. An early release for evaluation; not yet recommended for critical workloads.
The problem
A model nobody can explain is a model nobody is allowed to use.
Explanations are bolted on after training, in a notebook that only one person can run.
The path from experiment to a deployed, monitored model crosses three tools and loses the lineage on the way.
Risk and compliance teams are asked to approve a model from a slide, not from evidence.
Core capabilities
What MLDoo does.
Capabilities marked Planned are on the way and not yet in the product.
Visual model pipelines
Design the whole flow — data, features, model, evaluation, explanation — on one canvas.
Algorithms
Tabular models built on scikit-learn, with parameters exposed rather than hidden.
Neural network layers
Assemble networks layer by layer with shape inference, trained with PyTorch when it is installed.
Training
Real training runs with live progress, on the server or on a remote worker with a GPU.
Validation
Flows are checked before they run, including for data leakage between training and test sets.
Evaluation
Metrics and diagnostic charts computed from the run itself.
Feature engineering
Prepare and encode features as visible steps of the flow, not hidden preprocessing.
Explainability
Global and per-prediction explanations, including SHAP where the model supports it.
Model comparison
Compare runs and models side by side, and see what changed between two versions of a flow.
Versioning
Datasets, flows and models are versioned, so a result can be traced to exactly what produced it.
Registry
Register models with their evidence and move them through review before they are used.
Deployment
Serve a model behind an authenticated endpoint for predictions and explanations.
Monitoring
Track what a deployment is asked and how it answers after it goes live.
Model lineage
A tamper-evident audit trail links each model to its data, flow and run.
How it works
From first connection to a running result.
Step 1: Bring versioned data
Import a dataset and profile it. Each version is kept, so a model always points at the data it learned from.
Step 2: Design the flow
Lay out preparation, the model and its evaluation visually; MLDoo validates the design before you spend a training run on it.
Step 3: Train and read the evidence
Watch the run live, then read its metrics and explanations — what drives the model overall and why it made one prediction.
Step 4: Register, deploy, monitor
Promote the model with its evidence attached, serve it behind a token-protected endpoint and keep watching it.
Example workflow
What it looks like in practice.
One lineage from the dataset version to the monitored deployment.
- Dataset
- Features
- Train
- Evaluate
- Explain
- Register
- Deploy
- Monitor
Part of DooHorizon
How MLDoo connects to the platform.
Nothing is simulated: every metric, chart and explanation is computed from a real run — and a method that does not apply to a model is shown as unsupported, never faked.
FlowDoo
Clean and integrate data in FlowDoo, then model it; scoring rows with a deployed model from a flow is planned.
About FlowDooVizzDoo
MLDoo shares VizzDoo's data engine and design system, so exploring and modelling use the same foundations.
About VizzDooAIDoo
Status: Coming SoonAIDoo is designed to call deployed models as tools and to reason over a prediction together with its explanation.
About AIDooGovernanceDoo
Status: Coming SoonThe registry, approvals and audit trail are designed to surface in GovernanceDoo as model governance.
About GovernanceDoo
Security and governance
Built to be trusted with real data.
Designed with enterprise security and governance requirements in mind. No certification is claimed that has not been obtained.
How DooHorizon approaches security- Each model carries a Model Passport: owner, training data, performance, features, limitations, risk rating and approval status.
- An audit trail with a hash chain makes changes to the record evident.
- Training code runs in a sandbox; deployments are reached only with their own bearer token.
- Explanations are computed, not generated — and marked unsupported where a method does not apply.
Use cases
What you can build with MLDoo.
Predict customer or credit risk
Build a model whose individual decisions can be explained to the person affected and the team accountable.
Forecast demand or workload
Compare candidate models on the same data and keep the evidence for the one you choose.
Give reviewers evidence, not slides
Hand risk and compliance a Model Passport with the data, metrics and limitations in one place.
Learn ML with the lid off
See every step of the pipeline, which makes MLDoo a practical way to teach and learn modelling.
The product
A look at MLDoo.
Model Passport
- Model name
- Churn risk
- Version
- v3
- Owner
- Customer Analytics
- Training dataset
- customers_2026_q2 · version 7
- Performance
- AUC 0.87 on the holdout set
- Features
- 14 — tenure, plan, tickets in 90 days…
- Explainability
- Global drivers and per-prediction attributions
- Limitations
- Not validated for accounts younger than 30 days
- Risk rating
- Medium
- Approval status
- Approved
- Deployment status
- Serving · monitored
Prefer the real thing? Open MLDoo — it is free to start.
FAQ
MLDoo: common questions.
Is MLDoo focused on explainability?
Yes. Explainability and interpretability are intended to be first-class product principles in MLDoo, from the design of a flow to the record kept for each model.
What does “Preview” mean for MLDoo?
MLDoo is an early release for evaluation. The studio runs real training, explanation and serving, but it is not yet recommended for critical workloads.
Does MLDoo guarantee that a model is explainable?
No tool can. MLDoo is designed for transparency: it computes established explanation methods, shows their limits, and tells you when a method does not apply.
Can I train on a GPU?
Yes. Runs can be sent to remote workers that have a GPU; the model is saved so that serving does not need one.
Start with MLDoo. Expand when you are ready.
Start with one use case. Expand across the ecosystem.