Guardian
Guardian is Protect AI's security tool that scans machine learning environments for risks, misconfigurations, and potential vulnerabilities. Think of it as an always-on security system for AI pipelines.
Objective
Design an experience that makes AI/ML model security understandable, trustworthy, and actionable for engineers and analysts.
- Scan models across formats (TensorFlow, PyTorch, ONNX, LLM formats) and surface risks clearly
- Enforce policies in developer workflows — local scans, CI/CD gates, pre-download checks
- Provide fast triage-to-remediation paths, auditability, and enterprise readiness
Target Personas
"Before I download or deploy a model, tell me if it's risky and how to fix it."
"Automate checks in pipelines without slowing delivery."
"Get reliable signal, reduce false positives, and prove compliance."
"Show evidence that models meet internal and regulatory standards."
UX Approach
- Discovery interviews, ecosystem audit (model hubs, CI tools), competitive teardown
- Concept sketches to low-fidelity wireframes for critical flows and information architecture
- Built Radius Design System (tokens, components, patterns) to unify app and docs
- High-fidelity prototypes with realistic data and iterative usability tests
- Close partnership with PM/Eng on MVP scope, success metrics, and roadmap
Key Flows
Results detail → Fix guidance → Approve or Quarantine
Pick template → Configure policy → Test run → Gate on severity
Create rule → Scope to environments → Review impact → Publish
Prioritized queue → Evidence panel → Assign → Track to resolution
Time-series of findings, SBOM export, policy drift report
Challenges & Solutions
Users span ML, Platform, and SecOps — jargon differs. Solved with role-based defaults, glossary tooltips, and progressive disclosure.
Users distrust unreliable results. Calibrated severity and confidence badges, grouped duplicates, clear provenance, and one-click risk acceptance with expiry.
Local scanner path for pre-commit checks, CI templates with runtime estimates, dry-run mode, and policy preview.
Policy Builder with rule cards, natural language assist, scoping, simulation, versioning, and diff on publish.
Table virtualization, column presets per persona, smart grouping, and finding detail drawer with tabs.
Attestations/SBOM export, immutable audit log, report templates, and clear lineage from policy to scan to action.
Radius Design System
Built alongside Guardian to unify the product surface and accelerate delivery.
Design tokens, light/dark themes, density scales, motion
Primitives (Button, Input, Tabs, Table) and complex patterns (Scan Result Card, Finding Detail Drawer, Diff Viewer, Policy Builder, Severity Badge)
Progressive disclosure, inline explainers, review & publish workflows, wizard setup, zero-state guidance
WCAG 2.2 AA contrast, keyboard focus order, live regions for scan status, reduced motion support
Outcomes
Users reported faster understanding of what's wrong and what to do
Local scanning + CI templates reduced setup time; policy preview decreased gate failures
Role-tailored defaults and better evidence improved trust across teams
Radius DS accelerated new features and ensured consistency across the platform