HavenForKids

A Children's Emotional Health & Wellness Platform Powered by AI

Live: Open HavenForKids · Parent Wellness Dashboard

HavenForKids is a digital health and wellness platform designed to support the emotional development of children aged 7–12. Through a warm, browser-based pixel-art village, children engage with AI wellness companions that help them process feelings, build emotional vocabulary, and develop healthy coping habits — all through natural conversation. A real-time parent wellness dashboard surfaces emotional patterns, sentiment trends, and early distress signals — without ever exposing conversation text.

HavenForKids sits at the intersection of preventive mental health, child development, and responsible AI — giving families an accessible, stigma-free tool for emotional wellbeing long before a clinical intervention is needed.

The Wellness Gap We're Closing

Emotional health is the foundation of a child's overall wellbeing — yet it is the most consistently under-served dimension of children's health. The numbers are stark:

1 in 5 children experiences a diagnosable mental health condition before age 18
The average delay between first symptoms and first professional support is 12 years
Most children lack the vocabulary, confidence, or trusted channel to express what they are feeling
Existing tools are either clinical instruments built for professionals, or consumer apps that ask children to self-diagnose before they are developmentally ready

HavenForKids takes a wellness-first approach. Rather than treating emotional health as a clinical problem to be solved, we treat it as a daily practice to be supported — the same way physical health apps support exercise and nutrition. Children don't talk about their feelings. They explore a cosy village, meet friendly companions, and feel genuinely heard. The wellness insights emerge naturally from those conversations.

Core Wellness Features

Feature	Wellness Benefit
AI wellness companions	Three companions (Sunny, Sage, Keeper) each specialising in a distinct emotional domain — anxiety, stress, and fear — providing targeted, age-appropriate support
Proactive check-ins	Companions initiate conversations every 2–3 minutes, modelling healthy emotional check-in habits and reducing the barrier for children to open up
Persistent emotional memory	Companions remember previous conversations using vector-embedded semantic memory, building a continuous, trusting relationship over time
Real-time sentiment tracking	Dual-layer keyword-based scoring monitors emotional tone across every conversation, surfacing trends before they become crises
70+ distress detectors	Early-warning system covering bullying (physical, verbal, social), school avoidance, self-harm ideation, and social isolation
Parent wellness dashboard	Gives parents pattern-level emotional health insights — conversation frequency, tone trends, companion engagement, distress alerts — with zero exposure of conversation text
Family Code access	Privacy-first parent authentication via a 6-character code — no accounts, no surveillance, no data sharing
Child-safe design	4-digit PIN login for ages 7–12, Grade 5 reading level responses, browser-based with no installation required
Privacy by design	No third-party analytics, no advertising SDKs, no data sharing — all data stays in your own Convex deployment

How HavenForKids Supports Wellness

For the Child — Building Emotional Fitness

The child opens a browser and enters a small, safe village. They choose a character, name themselves, and explore freely. When they walk close to a wellness companion, a conversation begins naturally. The child types in their own words. The companion responds warmly, validates what the child shares, asks gentle follow-up questions, and models healthy emotional processing — always in simple language, always without judgement or diagnosis.

Companion	Wellness Domain	Approach
Sunny	Anxiety & worry	Normalises worry by sharing small concerns of their own; builds confidence through gentle validation
Sage	Stress & overwhelm	Teaches grounding and mindfulness through calm, structured conversation
Keeper	Fear & uncertainty	Builds resilience by breaking large problems into small, manageable steps

Each companion maintains persistent emotional memory. When a child returns the next day, the companion picks up naturally from where they left off — reinforcing the sense of a trusted, ongoing relationship that is central to emotional wellness.

For the Parent — Informed Without Intrusive

Parents access a separate wellness dashboard using a six-character Family Code generated at registration. The dashboard provides:

Which wellness companions the child has engaged with, and how often (interactive charts)
Real-time emotional tone scoring across conversations (1–5 scale, keyword-based NLP)
Context-aware wellness suggestions that adapt to detected emotional patterns
Early distress alerts when language associated with crisis, bullying, or school avoidance is detected
Engagement trends over time — last active timestamp, total conversations, weekly activity

The parent dashboard never displays conversation text. Parents receive wellness-level pattern insights, not surveillance. This boundary is enforced at the data architecture level — not just policy.

Safety & Crisis Response

Every message is evaluated by two independent wellness safety layers running in parallel:

AI-level analysis — Each companion is instructed to flag language indicating self-harm, abuse, or severe distress, and respond with warmth and safety
Keyword matching — A backend layer scans against 70+ distress phrases covering:
- Self-harm and suicidal ideation
- Physical bullying ("punched me", "kicked me", "hit me")
- Verbal and social bullying ("calls me names", "making fun of me", "laughed at me")
- School avoidance ("don't want to go to school", "hate school", "scared of school")
- Isolation and loneliness ("nobody likes me", "have no friends", "feel alone")

If either layer triggers, the parent dashboard immediately surfaces a distress alert. The child's experience is never interrupted or alarmed.

Platform Architecture

           Child Browser                    Parent Browser
                │                                │
    ┌───────────────────────┐       ┌──────────────────────────┐
    │  Wellness Game (Vite) │       │  Parent Dashboard (Vite) │
    │  Google Cloud Run     │       │  Google Cloud Run        │
    │                       │       │                          │
    │  PixiJS village       │       │  Recharts visualisation  │
    │  Three AI companions  │       │  Framer Motion UI        │
    │  Username + PIN auth  │       │  Family Code access      │
    └───────────┬───────────┘       └────────────┬─────────────┘
                │                                │
                └───────────────┬────────────────┘
                                │
               ┌────────────────────────────────┐
               │   Convex Serverless Backend    │
               │   Real-time subscriptions      │
               │                                │
               │   havenUsers    (accounts)     │
               │   worlds        (game state)   │
               │   messages      (chat log)     │
               │   memories      (AI memory)    │
               │   archivedConversations        │
               └────────────────┬───────────────┘
                                │
               ┌────────────────────────────────┐
               │   LLM Provider (Mistral AI)    │
               │   Companion conversations      │
               │   Memory summarisation         │
               │   Wellness reflection          │
               └────────────────────────────────┘

Key Design Decisions

Decision	Wellness Rationale
Username as stable identity	Agent memories are permanently tied to a specific child — companions build a genuine, continuous relationship rather than starting fresh each session
Family Code, not username	Parents access wellness insights without being able to identify or surveil individual conversations
Memory isolation	Each child has a completely separate memory graph — siblings or classmates using the same deployment never share emotional context
Proactive engagement	Companions initiate check-ins every 2–3 minutes, modelling the healthy habit of regular emotional check-ins rather than waiting for a child to reach out in distress
Conversation-aware idle protection	The idle-kick timer is suspended during active conversations — a child is never disconnected mid-disclosure
Agent leave guard	Companions cannot leave a conversation while a child is talking — reinforcing the safety and reliability of the relationship
Real sentiment scoring	Tone analysis uses genuine keyword-based NLP rather than placeholder values — wellness insights are grounded in real data

Technology Stack

Component	Technology
Wellness game engine	AI Town (Convex-native pixel-art village simulation)
Rendering	PixiJS
Game UI	React 18, Vite 6, Tailwind CSS
Dashboard UI	React 18, Vite 6, Framer Motion, Recharts, Lucide
Backend	Convex (serverless TypeScript functions, real-time subscriptions)
Database	Convex integrated database (indexed tables, vector search)
LLM provider	Mistral AI (mistral-large-latest) via OpenAI-compatible SDK
AI memory	Vector embeddings with semantic search (Convex vector store)
Child auth	Custom username + 4-digit PIN system
Wellness analytics	Dual-layer keyword-based scoring (70+ distress phrases, positive/negative word lists)
Game hosting	Google Cloud Run
Dashboard hosting	Google Cloud Run (Dockerised)

Live Deployments

Service	Platform	URL
Wellness Game	Google Cloud Run	havenforkids-game-708818914032.us-central1.run.app
Parent Dashboard	Google Cloud Run	havenforkids-dashboard-708818914032.us-central1.run.app
Convex Backend	Convex Cloud	`impartial-goldfinch-622.convex.cloud`

Wellness Companion Design

Each companion is defined by a unique identity, an emotional wellness archetype, and a set of behavioural constraints that prioritise the child's safety and development. Key principles enforced across all companions:

Responses are limited to 2–3 sentences in simple language appropriate for ages 7–12
Companions validate feelings before offering any guidance — emotional validation is the first step in healthy emotional processing
Companions never diagnose, never refer to professional help unprompted, and never dismiss what a child shares
When companions reference shared awareness of a child, they frame it as natural intuition ("I've been thinking about that kind of feeling lately") — preserving the child's sense of privacy while enabling coordinated wellness support
Companions never leave a conversation with a child — message limits and duration timeouts only apply to agent-to-agent conversations

Privacy & Ethical Wellness Design

Privacy First. The parent dashboard provides pattern-level wellness visibility — frequency, tone score, distress flags — and does not expose any conversation transcripts. No third-party analytics, advertising SDKs, or tracking of any kind. Family Codes (not usernames) are the parent access mechanism. This is enforced at the data architecture level.

Do No Harm. Companion prompts explicitly prohibit medical advice, diagnosis, and unsolicited referrals. A dual-layer distress detection system (AI flag + 70+ keyword matching) alerts parents when crisis-associated language appears — without interrupting or alarming the child.

Inclusive Wellness Design. Four-digit PIN authentication accessible to young children. Visible Send button alongside keyboard submission. Grade 5 reading level for all companion responses. Browser-based — no installation, no app store, no device restrictions.

Transparency. Companions are presented as storybook characters, never as real beings or therapists. System prompts instruct companions never to claim personhood or professional expertise. This documentation describes the full technical architecture openly.

Zero Token Waste. AI calls are gated behind human-presence detection — no compute is consumed when no child is actively using the platform.

Acknowledgements

HavenForKids is built on top of AI Town, an open-source framework for simulating AI agent worlds by a16z-infra (MIT License). The world simulation engine, agent tick loop, pathfinding, and memory infrastructure are derived from that work.

HavenForKids' original contributions include: the wellness-focused authentication system, companion identities with ethically constrained prompts, human-presence-gated token management, proactive wellness check-in behaviour, conversation-aware idle protection, agent leave guards, shared companion memory with intuition rewrites, keyword-based sentiment analysis, 70+ phrase distress detection, and the parent wellness dashboard.

Additional credits: