Product flow
The AI System record, from first classification to reviewer handoff.
Epok is organized around the operational objects model teams already use: systems, model versions, datasets, logs, AutoML runs, required fields, and generated documents.
The core product idea is intentionally simple: regulatory evidence should be assembled from the same technical trail that produced the model. A model version should know which datasets, training runs, AutoML artifacts, evaluation metrics, deployment assumptions, and review decisions support it.
Epok keeps those objects connected so a reviewer can see what was captured automatically, what was generated from deterministic templates or AutoML, and what still requires human judgement.
Evidence package
Customer support classifier
Readiness
AI System
27 fields
Registry
Model Version
v2.4.1
Data governance
Dataset Card
Article 10
Generated evidence
AutoML Run
5 artifacts
Classify the AI system
Start with intended use, deployment context, users, outputs, and risk rationale. Epok turns that into an AI System Record with evidence requirements.
Generate evidence with AutoML
For tabular systems, run project-scoped AutoML to produce leakage reports, Dataset Card updates, Model Versions, and Evidence Logs.
Generate deterministic drafts
Draft documents come from templates and stored project evidence, keeping the record auditable instead of LLM-written from scratch.
Review and export
Reviewers see captured, generated, and review-required evidence before anything becomes a review package.
Evidence graph
A product record that can explain itself.
Instead of asking teams to recreate context in a late compliance document, Epok links every draft section back to source objects. The graph is not decorative. It is how teams answer: where did this statement come from, who reviewed it, and what is still missing?
App cards
Screenshot-style surfaces for the evidence reviewers ask about.
Readiness
AI System
27 fields
Registry
Model Version
v2.4.1
Data governance
Dataset Card
Article 10
Generated evidence
AutoML Run
5 artifacts