MindCare AI: Multi-Label Emotion Detection from Text for Mental-Wellbeing Awareness with English, Hindi and Marathi Support

Emotional distress often goes unnoticed, and text tools that turn what people write into emotion signals could support self-awareness, yet many such tools are English-only, give a single score, and say little about their errors, privacy or limits. This paper presents MindCare AI, an end-to-end prototype that predicts multi-label emotions (27 emotions plus neutral) from short text. A TF-IDF (1–2-gram) representation feeds 28 one-vs-rest logistic regressions trained on the official GoEmotions splits (43,410 / 5,426 / 5,427 comments), with a single decision threshold (0.45) chosen on validation data only. On the official test split the model reaches macro-F1 0.445, micro-F1 0.514, micro precision 0.430, micro recall 0.638, subset accuracy 0.276 and Hamming loss 0.0503; the top-scoring label is one of the true labels for 56.9% of comments. Per-label F1 ranges from 0.888 (gratitude) to 0.129 (relief), and eight of 28 labels stay below 0.30. The model over-predicts (1.73 labels per comment against 1.17 annotated) and assigns neutral to 31.6% of comments whose annotation is not neutral. The classifier is embedded in a Flask web application with registration, opt-in text storage, per-user history, a JSON API, an English / Hindi / Marathi interface with a small phrase-and-emoji normalisation lexicon, rule-based suggestions, a 10-item self-check questionnaire and a keyword-based crisis notice. These last four components have not been validated, and transformer baselines were not run. This is a single-seed study of a prototype; it is an awareness aid, not a diagnostic tool.