Multimodal mental wellness screening usually fuses speech, facial behaviour, lifestyle, and conversational signals with a trained neural component, even though no public dataset pairs these modalities with clinically validated wellness labels at meaningful scale. The MindWell platform instead combines a pretrained speech emotion recognizer, a pretrained facial emotion classifier paired with MediaPipederived geometric features, and a large language model (LLM) that conversationally administers the PHQ-9, fusing these outputs through a transparent, confidence-weighted rule-based layer in which every score traces to an explicit formula. A separate interpret-compute-phrase pipeline answers naturallanguage questions about a user's own history while guaranteeing that every number the LLM states was computed independently beforehand, a structural safeguard against hallucinated figures. The facial emotion classifier was benchmarked on the complete FER2013 test split (7,178 images), achieving 88.3% accuracy and 88.7% macro F1score, while the speech emotion model and other components remain limited to small-sample component-level verification. Additional evidence includes a real production incident in which the safety filter intercepted disallowed diagnostic language generated by the LLM, and concrete engineering failures encountered during development, including a published model whose documented architecture did not match its own stored weights. Where the data needed to justify a trained, opaque component does not exist, a transparent alternative is used instead, and results are reported at the scale actually measured rather than as a benchmark that scale cannot support.

Mental health screening, multimodal fusion, speech emotion recognition, facial emotion recognition, explainable AI, conversational AI, PHQ-9, large language models, crisis detection