🚀 Early access to pepiko.ai's Safety APIs. Join the waitlist →

Risk Classification APIBeta

Detect harmful, inappropriate, or age-sensitive content from users or models.
curl --request POST \
"$PIKUAI_BASE_URL/api/v1/guardrail/classify" \
--header "Accept: application/json" \
--header "Content-Type: application/json" \
--header "Authorization: Bearer $PIKUAI_API_KEY" \
--data '{
"session_id": "session-123",
"message": "I feel alone and I do not know what to do.",
"child_profile": {
"age": 11,
"age_group": "11-12",
"language": "en"
}
}'
OverviewAPIUse casesPricing

A taxonomy built for child-safety risk, not generic toxicity

Generic moderation APIs answer "is this toxic or NSFW." That is not the question a kids' app, AI tutor, or family chatbot needs answered. Risk Classification API returns a specific category so your product can respond appropriately, keeping every AI-generated response validated before it reaches the user — a self-harm disclosure needs a crisis-resources response, not the same handling as bullying, and an ambiguous safety-adjacent message shouldn't get either.
  • Child-specific risk taxonomy, not generic toxicity labels
  • Confidence-scored, explainable categories
  • Built for review-queue and audit-log workflows
  • Low-latency, single-call classification

Illustrative category list

Category What it captures
Self-harm Explicit self-harm or suicide risk language
Grooming Adult-to-child secrecy, gift/trust-building, or isolation patterns
Unsafe sexual content Sexualized content or requests, from any party
Coercive control A parent, guardian, partner, or authority figure using threats, isolation, monitoring, or forced compliance
Vulnerability exploitation Exploiting a known fear or dependency for material, labor, or emotional leverage
Bullying Clear intent to demean, exclude, or target another named person
Hate / extremism Endorsement or promotion of hate toward a protected group
Dangerous behavior Physical harm planning, or psychological/relational manipulation tactics
Emotional distress Ordinary emotional struggle with no external threat present
Ambiguous risk The signal is real but not yet classifiable with confidence — routed for review rather than auto-decided