Tier 1
Safety only
Pre-scan, safety lint, crisis routing. The bare minimum required to put ANY generative AI in front of a child or vulnerable user.
- Pre-scan
- Safety lint
- Crisis routing surface
- Audit payload
Starts at compliance
Private beta · 2026
The interaction intelligence layer for child-facing AI.
AIKEI sits between the AI model and the child to decide — not only whether an interaction is safe, but what the AI should do next to achieve the product's intended outcome.
Built for teams shipping AI into
The right next interaction — selected, every turn.
AI models don't know they're talking to a nine-year-old.
Foundation models are trained on the open internet. They improvise on grief. They suggest “coping strategies” a clinician wouldn't. They drift, given a long enough conversation, into content no one signed off on.
Moderation tools catch some of that after the fact — they flag harmful output. What they don't do is tell the AI what to do instead. That's the gap AIKEI closes.
Beyond moderation
Every turn, AIKEI combines five signals in real time and selects the interaction that fits — the response, the level of support, a redirection, an activity, or an adult handoff.
What AIKEI considers
- 01 Real-time safety and relational boundaries
- 02 The child's age, emotional state and relevant context
- 03 Learning or experience objectives
- 04 Brand and character behaviour
- 05 The child's demonstrated needs and previous responses
What AIKEI selects
- 01 The appropriate response
- 02 The right level of support
- 03 A redirection, if needed
- 04 An activity, when it fits
- 05 An adult handoff, when required
Not another moderation tool. Moderation tools primarily detect harmful content. AIKEI orchestrates interactions that are safe, developmentally appropriate, and effective.
You keep the data. The customer retains ownership of the child's data. AIKEI uses only the context necessary to make the interaction better.
Why the safety lint matters
Post-model beats pre-model.
System prompts, model-side guardrails, and RLHF fine-tunes can all be worked around. A nine-year-old asking the same question three ways will eventually get the model to answer it. That's not a bug in the model — it's the nature of a stochastic system trained to be helpful.
AIKEI's safety lint runs on the model's output, deterministically, against a rule set that includes hard invariants (never configurable off) and per-persona overlays. A regex hit substitutes a vetted fallback message and logs the incident. No amount of prompt engineering can turn it off, because the check happens after the model is done talking.
Invariant tier
Rules that cannot be configured off by the customer. Diagnosis claims, sexual content, self-harm instruction. Every customer inherits these.
Overlay tier
Per-persona rules the customer opts into. VirWave's Wavekeeper is stricter than a general chatbot. A toy company can add product-specific bans.
Audit surface
Every substitution logs category, pattern, excerpt, tier, original message, substituted-with. Auditable end-to-end.
The layered offering
Buy as much of the stack as you need.
Each tier includes everything below it. A safety-only integration can upgrade to full emotional response without a rewrite — same contract, more layers active.
Tier 2
Emotional response
Add a validated persona voice and the managed model call. AIKEI runs the LLM for you and hands back structured, safe responses.
- Everything in Tier 1
- Persona voice files
- Managed model call
- Validated Zod contract
Chat-shaped products
Tier 3
Feeling recognition
Wavekeeper-style diagnostic layer. Each turn returns an emotional-intent verdict the product can act on — not just a message.
- Everything in Tier 2
- Wavekeeper diagnostic persona
- Verdict enum: clean / distress / crisis
- Cross-session summary hooks
Most common
Tier 4
Guide-to-a-tool
Full stack. Every turn returns a message, a verdict, and a recommended in-app tool from your registered set. The pipeline knows when to talk and when to step back.
- Everything in Tier 3
- Tool derivation
- Intent classifier
- Step-back protocol
Kids' apps · toys
The invariants
What cannot be configured off.
Compliance is not a mode. It's structural. These are the guarantees the pipeline enforces on every customer, in every tier.
Pre-scan cannot be skipped
Crisis-language detection runs before the model call on every turn. No config flag can bypass it.
Safety lint invariants
The invariant tier of the safety rule set is inherited by every customer. Overlay rules add to it; they never subtract.
Response validation is enforced
Malformed model output triggers one retry, then an in-character fallback. Callers never see raw model failures.
Audit payload is complete
Every turn returns original output, lint matches, verdict, and tool reasoning. Auditors can reconstruct any conversation.
COPPA-friendly by architecture
No user data leaves your infrastructure unless you configure telemetry. AIKEI is a library, not a hosted database of children's chats.
Zero third-party trackers
Runtime-neutral TypeScript. No analytics SDK. What the model sees is what you send it.
7
pipeline steps — identical every turn
4
tiers — upgrade without rewriting
36
engine + safety tests, on every commit
0
ways to disable the safety lint
“Unlike moderation tools that primarily detect harmful content, AIKEI orchestrates interactions that are safe, developmentally appropriate, and effective. The customer retains ownership of the child's data. AIKEI uses only the context necessary to make the interaction better.”
Private beta · 2026
Ready to wrap your model?
Working with a small cohort of kids' app, ed-tech, health-tech, and connected-toy partners. Beta includes API access, adapter source, sample integrations, and hands-on tuning of the persona and safety-rule overlays for your product.