Rescue plan AI

Your mind matters – plan before the storm

Comment

Inspiration

Mental health crises often happen without warning – especially among students. We discovered that 58.6% of students in a public dataset of 27,901 students showed signs of depression. Yet most students don't have a plan for when things get hard. They wait until crisis.

We wanted to change that. We built RescuePlan AI – a tool that predicts depression risk using machine learning and generates a personalised safety plan to help students prepare before they need it.

What it does

AI Risk Assessment: Users answer 10 simple questions (age, academic pressure, sleep, suicidal thoughts, etc.). Our model predicts their depression risk score (0-100%) with 84% accuracy and 0.92 AUC.

  • Personalised Safety Plan: Based on their answers, the app auto‑generates a 6‑section safety plan:
    1. Warning Signs
    2. Coping Strategies
    3. Supportive People/Places
    4. People to Ask for Help
    5. Professional Contacts
    6. Safer Environment
  • Risk Breakdown: A visual chart shows why they got their score – each factor's contribution.
  • Dashboard: Aggregated insights from the dataset (city‑wise depression rates, top predictors).
  • Resource Library: Curated mental health resources based on risk level.
  • Bilingual: English and Tamil support.
  • Privacy‑first: No accounts, no data storage. ## How we built it Frontend: Streamlit (Python) – fast, interactive, and easy to deploy.
  • Backend: Scikit‑learn for ML (Logistic Regression).
  • Data: Public Kaggle dataset of 27,901 students (18 features).
  • Model Training: Preprocessing (encoding, imputation, scaling) → training → evaluation (Accuracy 84.4%, AUC 0.92).
  • Visualisation: Matplotlib for charts.
  • Deployment: Streamlit Cloud (free, public URL). ## Challenges we ran into
  • Handling missing data: Used SimpleImputer with median strategy.
  • Model interpretability: Made SHAP‑like contribution charts so users understand their score.
  • Deployment on Streamlit Cloud: Had to manage dependencies (requirements.txt) and ensure the dataset loaded correctly.
  • Balancing features: The model needed to be accurate but simple enough to deploy quickly.

Accomplishments that we're proud of

Built a complete, live, interactive web app in a short time.

  • Achieved 84% accuracy and 0.92 AUC – comparable to production models.
  • Made it bilingual (English + Tamil) and accessible.
  • Designed a privacy‑first tool – no data stored, no accounts required.
  • Created a dashboard and resource library to go beyond just a predictor ## What we learned
  • Logistic Regression can be surprisingly powerful for mental health screening.
  • Interpretability is critical – users (and judges!) trust models they understand.
  • A good UI makes a huge difference in user engagement.
  • Streamlit is incredibly fast for prototyping and deploying ML apps. ## What's next for Rescue plan AI
  • Multi‑language expansion: Add more Indian languages.
  • Counsellor dashboard: Let universities see aggregate risk trends (anonymised).
  • Mobile app: React Native version for offline access.
  • Integration with helplines: Real‑time crisis button.
  • More data: Incorporate mood‑tracking journals to improve predictions.

Built With

Share this project:

Updates