Design AI safety systems.

### Signal to interviewer

I can design safety systems as an integrated architecture with measurable controls, not isolated moderation endpoints.

### Clarify

I would clarify risk domains, policy boundaries, legal requirements, and acceptable false-positive tolerance.

### Approach

Implement a defense-in-depth safety mesh: pre-input screening, model-time policy conditioning, post-output moderation, and escalation pathways.

### Metrics & instrumentation

Primary metric: prevention rate for high-severity harms. Secondary metrics: false-positive moderation rate, escalation resolution time, and policy consistency score. Guardrails: overblocking, bypass attempts, and safety regression recurrence.

### Tradeoffs

Stronger restrictions increase safety but can reduce utility for benign edge cases. Faster responses improve UX but may weaken safety review depth.

### Risks & mitigations

Risk: adversarial prompt evolution; mitigate with continuous red teaming. Risk: policy drift across features; mitigate with centralized policy service. Risk: opaque user experience; mitigate with transparent refusal rationale.

### Example

For a public assistant, health-related responses require grounded sources and trigger human escalation when confidence is low and potential harm is high.

### 90-second version

Design safety as layered controls across the full request lifecycle. Measure severe harm prevention, tune by risk tier, and continuously update defenses against evolving misuse.