Project: Confidex Analytics — Computer Vision-Driven Behavioral Scoring Pipeline

Developed a high-accuracy, research-backed computer vision and deep learning pipeline designed to analyze non-verbal upper-body ergonomics and hand kinematics from video streams to calculate objective confidence metrics.

Key Technical Implementations: • Landmark Feature Extraction: Built a real-time frame processing pipeline using OpenCV and MediaPipe to convert raw interview videos into structured sequential tracking data, capturing spatial coordinates of physical gestures and skeletal movements from 200+ video datasets. • Deep Learning Architecture: Evaluated and compared recurrent neural networks (LSTM, GRU, BiLSTM) to model temporal gesture development across frame streams, selecting and optimizing a high-performance BiLSTM sequence model that achieved a verified 82.79% prediction accuracy. • Temporal Modeling & Softmax Aggregation: Preserved the model's entire uncertainty distribution layer by retaining complete softmax outputs. Structured a temporal pipeline that aggregates these class probabilities across strict 1-second rolling window frames to maintain maximum representation stability. • Research-Based Weighted Fusion: Engineered a quantitative scoring engine that maps normalized gesture frequencies to research-derived behavioral coefficients (Hands on table: 0.322, Hidden hands: 0.264, Gestures on table: 0.207, Self-touch: 0.207). Programmed dynamic inversions for negative behavior inputs (Score = 10 - Negative) to output a clean, unified confidence rating out of 10. • Backend Deployment: Exposed the finalized end-to-end deep learning scoring pipeline via a high-performance FastAPI web service layer for production readiness.