Sentinel — Privacy-First Behavioral Safety Agent

With 1 in 7 teenagers contacted by online predators and 4 in 10 experiencing cyberbullying, existing parental control tools (Bark, mSpy) rely on invasive message scraping or screenshots that break teen trust and fail when platforms restrict third-party APIs.

How to detect complex, evolving behavioral risks (grooming, harassment, addiction) early while preserving teenager privacy, avoiding rigid rule-based false positives, and keeping parents in the loop for critical decisions?

First-party OS behavioral metadata (iOS Screen Time DeviceActivity API, Android UsageStatsManager, notification timing): daily screen time, per-app session frequency, nocturnal activity drift (10pm–6am), inferred sleep windows, and anonymized per-contact interaction cadence and response latency across a 14-day rolling baseline and 7-day analysis window.

Architected a 4-node LangGraph multi-agent workflow sharing a typed MetadataWindow state: (1) Collector (deterministic validation, no LLM), (2) Analyser (GPT-4o evaluating 4 specialized skills: Grooming, Harassment, Addiction, and Normal Life Event disambiguation such as exams or new healthy friendships), (3) Scorer (GPT-4o correlating multi-signal patterns against the teen's personal 14-day baseline on a 0–1 risk scale), and (4) Communicator (GPT-4o-mini). Enforced architectural guardrails: human-in-the-loop parent escalation only above 0.65, full chain-of-thought audit logging, Trusted Contacts exclusion, Pause Mode, and feedback-driven baseline recalibration.

Delivered a functional multi-agent prototype capable of distinguishing genuine multi-signal risk escalations (e.g., nocturnal drift + platform switch + engagement asymmetry) from benign exam spikes (capped below 0.38), with full reasoning traceability and zero message-content inspection.

Showcases responsible agentic AI architecture — combining LangGraph state orchestration, least-privilege guardrails, human-in-the-loop governance, and privacy-by-design.