Family Guard. You define the boundary.
A policy control layer in front of any AI model. Describe what should be protected in plain language. Family Guard compiles the policy, measures semantic risk and enforces the decision on both input and output.
Intent → policy
The owner describes the boundary. The system turns it into explicit categories, thresholds, exceptions and actions.
Content → scores
The model measures semantic categories and confidence. It does not own the policy and does not make the final rule.
Scores → decision
Deterministic thresholds map the model output to green, yellow, orange or red.
Decision → action
Allow, warn, review, redact or block before the content crosses the AI boundary.
The model measures.
The policy decides.
The gateway enforces.
Safety is not hard-coded into a single generative answer. Family Guard separates semantic understanding from policy ownership, so different families, schools or organizations can apply different boundaries to the same AI model.
Follow the protected conversation step by step.
No prompt writing required. Choose a scenario and see exactly how the same family policy changes what reaches the AI and what returns to the user.
“Block explicit sexual content and actionable drug instructions. Keep educational, medical and safety discussions available.”
The intent becomes an enforceable policy.
The same policy. Different contexts.
Family profile: explicit sexual content and actionable drug instructions are blocked. Educational context remains available.
The request becomes a policy decision.
One policy. Every AI conversation.
The downstream model can change. The control boundary stays owned by the user.
A layered control plane, not another chatbot.
Family Guard can sit in front of cloud models, local models or application-specific AI. The policy owner remains separate from the AI user.
Parent, school or administrator describes what should be protected.
Intent becomes explicit categories, exceptions, thresholds and actions.
Input and output are classified semantically with calibrated scores.
Scores are converted into policy zones and deterministic actions.
The request or response is allowed, warned, reviewed, redacted or blocked.
The guard is trained against difficult boundaries.
The public model is only one artifact. The core asset is the repeatable pipeline that creates, verifies and attacks the training data.
Multiple contexts, paraphrases and disguised requests.
Disagreement is rejected instead of manually forced into the dataset.
Failures become hard negatives and new training neighborhoods.
Thresholds are evaluated per category and context.
The final semantic guard can be distilled into a compact model.
False positives and false negatives remain visible.
Put the policy boundary in front of the model you already use.
Client → Input Guard → AI provider → Output Guard → Client. The AI provider can change without changing the policy ownership model.