SkinAgent AI is a consumer‑facing, non‑diagnostic skincare assistant that coordinates visual evidence, product knowledge, tool usage, and user actions within explicit evidence and safety boundaries. The architecture consists of three routing branches—Acne, Pores, Wrinkles—paired with photograph‑based skin‑type estimation and count‑informed ordinal acne‑severity support. Each branch invokes typed tools (e.g., product lookup, usage guide generation) whose outputs are grounded in a curated database. A deterministic safety layer enforces privacy filtering, evidence verification, approval before any state‑changing action, and provides structured trace‑and‑replay capabilities.
Evaluation was split into visual‑model performance and system‑level behavior. Across three random seeds the routing model achieved 99.84% ± 0.07% accuracy. Skin‑type estimation reached 88.85% accuracy, while the count‑informed acne‑severity module obtained 84.59% accuracy with a quadratic weighted kappa of 0.9076. On a locked, non‑independent benchmark of 240 cases, intent accuracy was 80.00%, exact tool‑set match 62.92%, and strict task completion 47.08%. No violations or cross‑user leakage were observed in the finite safety and privacy test suites.
Remaining challenges include tool‑selection errors, incomplete grounding of product attributes, and unreliable failure fallbacks. These results demonstrate the feasibility of bounded, database‑grounded, and traceable agent orchestration for non‑diagnostic skincare assistance, yet they do not establish clinical readiness, external generalization, formal privacy guarantees, or universal safety. Independent validation, expert assessment, robustness and fairness testing, and prospective real‑world evaluation remain necessary.
Review