Adrian: An Embodied, Privacy‑First Platform for Ethical Human–AI Interaction

Abstract

Modern artificial intelligence (AI) often elevates speed and engagement above privacy, transparency, and autonomy. Adrian proposes a different center of gravity: human dignity and reflective presence. The companion operates entirely on edge computing hardware with a modular, local-first software stack; background recording is disabled, system state is surfaced through clear indicators, and any data retention remains explicitly consent-gated. The manuscript details architecture for on-device speech input and output and lightweight perception, engineering trade-offs of local inference, and a replicable build process suitable for students and hobbyists in open-source robotics. An evaluation plan emphasizes human AI interaction signals such as comfort, perceived respect, and comprehension of data flows alongside throughput and latency checks. Contributions include: (i) a reproducible reference design for privacy preserving embodied AI that avoids dependence on cloud telemetry, and (ii) an interaction model that treats deliberate pacing and constrained memory as features that strengthen user agency and trust rather than defects to be engineered away. A short roadmap outlines public release, documentation, and extensions to additional on-device modalities. Quiet by design and precise in scope, Adrian argues that usefulness can arrive without surveillance, and that technology serves best when presence is measured, transparent, and owned by the human.