Back Posture Device

I was inspired to build this tool because my grandfather suffered from chronic back pain, and I wanted to create something that could help people prevent posture-related issues before they become serious.

While working on the project, I learned how to use machine learning, pose-estimation, and data preprocessing to build a reliable classification system. I also learned how small variations in angles, lighting, and body shapes can dramatically affect model performance, and how to fine-tune a model using metrics like accuracy, loss curves, and validation splits.

To build the system, I used pose-estimation keypoints as input features and trained a classifier to distinguish between good and bad posture. After experimenting with feature engineering and multiple model architectures, I achieved about 95% accuracy. The final system provides real-time posture feedback and is now used in 10 centers across India.

One challenge I faced was collecting clean, diverse training data. Another was making the model robust to camera differences and real-world environments outside controlled lab conditions. Tackling these issues helped me understand the importance of dataset quality and iterative testing.

Built With

• java
• javascript
• tailwind