Projects
MVP foundation

GuardianLLM

Age-aware AI safety middleware for child-facing LLM applications.

System flow

How the pieces move.

Problem Statement

Child-facing AI products need developmental safety, privacy protection, and explainable policy decisions before raw prompts and responses move through provider calls.

Architecture Summary

GuardianLLM runs a modular safety pipeline for age profiling, PII redaction, topic and intent classification, emotion detection, risk scoring, policy selection, provider calls, and output validation.

Key Design Tradeoffs

  • Starts with deterministic, inspectable rules instead of opaque classifiers.
  • Adds safety checks before and after provider calls, which is more deliberate than direct routing.
  • Treats provider safeguards as one layer, not the whole child-safety strategy.

Technologies

PythonFastAPIPydanticOpenAI adapterSafety policy

GuardianLLM explores how child-facing AI products can route model interactions through an age-aware safety layer before and after provider calls.

The MVP focuses on an inspectable foundation: age profiles, privacy redaction, topic and intent classification, emotion detection, risk scoring, policy decisions, output validation, API endpoints, SDK usage, tests, and a starter ChildSafeBench evaluation suite.

The project is intentionally not a production child-safety product. It is a reference implementation for the architecture around the product: the privacy boundary, policy decision, provider abstraction, validation pass, and metadata that make safety behavior easier to inspect and improve.