Building AI that
actually works.

We introduce VietCheckMed, the first framework for Vietnamese explainable, evidence-grounded regulatory compliance checking, alongside VietAestheticAds — the first large-scale, expert-validated benchmark comprising 8,329 advertisements paired with an authoritative corpus of 9,978 registered facilities. The task is fundamentally distinct from traditional fact-checking: rather than assessing veracity, it verifies legal authorization. We demonstrate that evidence-grounded RAG is essential, outperforming powerful unassisted LLM baselines by over 0.3805 F1. Primary remaining challenges are nuanced failures in semantic and logical reasoning, defining a clear frontier for future work.

Automated oversight of medical and aesthetic advertising requires authorization-based compliance checking rather than conventional truthfulness verification. The problem is especially challenging under cross-modal regulatory evasion, where benign or euphemistic text is used to conceal images of unauthorized procedures. Existing vision-language models are brittle in this setting: they often over-rely on the apparently safe modality, exploit spurious correlations, and require retraining when regulations change. We present MM-RegCheck, a training-free framework for multimodal regulatory compliance checking. It first extracts and grounds potential clinical procedures from image-text pairs in retrieved statutory evidence. To enforce compliance, we introduce TF-BEAR, an inference-time causal adjudicator. By comparing prior and posterior label odds, TF-BEAR isolates the specific contribution of the regulatory evidence, allowing the system to enforce strict liability when hidden visual evidence contradicts the text. We also introduce MedReg-MM, an expert-validated benchmark of real-world ads featuring adversarial cross-modal evasion subsets. With a 9B backbone, MM-RegCheck achieves 0.8985 Macro-F1 in Beauty and 0.6556 in Healthcare, attains the best explanation quality on FIDES (3.7), and transfers effectively to text-only regulatory checking. These results suggest that training-free causal reasoning is a practical and adaptable foundation for high-stakes multimodal compliance monitoring.

We introduce the first methodological framework for detecting legal violations in Vietnamese beauty and aesthetic industry advertisements on social media. Illegal ads include unlicensed businesses, wrong-location operators, and false claims not matching government registrations. We contribute a novel dataset of 1,175 Facebook advertisement posts from HCMC beauty businesses paired with their official government-registered information. Our integrated approach combines linguistic feature extraction with semantic content matching, achieving accuracy 0.783, precision 0.686, and F1 0.703 with the BGEM3 model — establishing the foundation for the VietCheckMed system.