Introducing Horizon: Child Safety Detection That Never Leaves the Phone

Introducing Horizon: Child Safety Detection That Never Leaves the Phone

Every parent faces the same impossible choice: monitor your child's messages and invade their privacy, or trust blindly and hope for the best. Existing safety tools force this tradeoff — they either upload private conversations to remote servers for analysis, or rely on simple keyword filters that miss nuanced threats and flag innocent messages.

Today we're introducing Horizon, SafeCircle's on-device AI model that changes this equation entirely. It detects real safety risks — grooming, bullying, threats, and more — in under 100 milliseconds, directly on your child's phone. No messages are ever sent to our servers. No one reads their conversations. Not even us.

The Problem With Current Solutions

Server-side content moderation works by sending your child's messages to a company's cloud infrastructure, where AI models scan them for risks. This approach has two fundamental flaws:

1. Privacy violation by design — every message passes through a third party's servers, creating a permanent record of your child's private conversations.
2. Single point of failure — if the service goes down, monitoring stops entirely.

Keyword-based filtering is the alternative, but it's trivially circumvented ("meet me at the p4rk") and generates endless false positives from normal teenage conversation.

How Horizon Works

Horizon takes a fundamentally different approach. Instead of sending data to the cloud, we bring the AI to the device.

The system uses two models working in tandem:

• Horizon-Full is our high-capacity teacher model — a 3 billion parameter language model fine-tuned specifically for risk detection. It understands context, subtext, and the subtle patterns that characterize real threats. This model runs on our servers during training only.
• Horizon-Mobile is a tiny 25MB classifier that runs entirely on-device. We trained it using *knowledge distillation* — a technique where a large, accurate model teaches a smaller model to replicate its decisions. The result is a mobile model that captures the intelligence of the larger system in a package small enough to run on any modern smartphone.

What It Detects

Horizon identifies seven categories of risk, each trained on thousands of realistic conversation examples:

• Grooming — incremental boundary violations, secrecy demands, trust manipulation
• Bullying — repeated intimidation, social exclusion, humiliation
• Sexual content — explicit or suggestive material directed at minors
• Isolation — attempts to separate a child from their support network
• Personal information — solicitation of addresses, schools, photos
• Platform migration — requests to move to less-monitored channels
• Threats — direct or implied physical or psychological harm

For each detection, Horizon provides a severity rating (none through critical), a confidence score, and — critically — an explanation of which specific words or phrases triggered the alert.

Privacy by Architecture

This isn't privacy by policy. It's privacy by architecture. The 25MB model file lives on the device. Inference happens on the device. Message content never touches a network connection.

There's no server to breach. No database of children's messages to leak. No employee who could access private conversations. The mathematical impossibility of extracting training data from a neural network means that even the model file itself reveals nothing about any individual child's messages.

The Technical Innovation

Building a model this small without sacrificing accuracy required solving several hard problems:

Knowledge distillation — We don't just compress the model; we teach the small model to replicate the reasoning patterns of the large model. When the teacher model says "this is 70% grooming, 20% isolation, 10% benign," that uncertainty itself is valuable training signal.

Dual-head classification — A single forward pass through the model simultaneously classifies both the risk category and severity level, keeping inference under our 100ms target.

Explainability — Using Integrated Gradients, we compute which specific tokens in the conversation most influenced the model's decision. Parents see why something was flagged, not just that it was.

What's Next

Horizon is currently trained on English-language conversations. We're actively expanding to Spanish, Portuguese, and French. We're also working on temporal pattern detection — identifying risk that emerges across multiple conversations over days or weeks, not just within a single message.

The full technical paper, including mathematical formulations, architecture details, and training methodology, is available for download below.

Full Paper