Solution
Understand why the model made the prediction.
Build models whose behaviour can be shown to the people accountable for them — with the evidence kept beside the model.
The problem
What teams are dealing with.
- A model performs well and still cannot be used, because nobody can explain a single decision to the person it affects.
- Explanations live in a notebook, disconnected from the model that is actually deployed.
- Reviewers receive a summary rather than the evidence.
Why it stays unsolved
Existing approaches fragment the workflow.
- Training, explanation, registry and serving sit in different tools, so the explanation and the deployed model can drift apart.
- Documentation is written by hand after the fact.
The DooHorizon solution
MLDoo puts interpretability at the center of the model lifecycle.
MLDoo
Status: PreviewComputes global and per-prediction explanations from real runs and keeps them with the model.
About MLDooVizzDoo
Status: Public BetaLets teams explore the data before modelling it.
About VizzDooGovernanceDoo
Status: Coming SoonIs designed to carry the Model Passport into review and approval.
About GovernanceDoo
Reference architecture
How the pieces are arranged.
- Versioned dataset
- MLDoo flow
- MetricsExplanations
- Model Passport
- Approved deployment
Example business workflow
Evidence first, then deployment.
- Train
- Evaluate
- Explain globally
- Explain one prediction
- Record limitations
- Approve
Business outcomes
What this lets you do.
- Explain an individual prediction to the person it concerns.
- Show reviewers what drives a model and where it is weak.
- Keep the explanation attached to the exact model version that is serving.
- Know when an explanation method does not apply, instead of trusting a chart that should not exist.
Governance and security
How it stays under control.
- No claim of perfect explainability: methods and their limits are shown as they are.
- A tamper-evident audit trail links model, data, flow and run.
- Deployments answer only to their own bearer token.
Start with this use case.
Start with one use case. Expand across the ecosystem.