I engineered a self-healing tabular pipeline that detects streaming distribution shifts and adapts without manual retraining. Its statistical monitors use KS tests and PSI, while TreeSHAP helps identify drifting features.
I designed a threshold-driven retraining loop that updates sliding windows when Wasserstein distance crosses a set threshold. During simulated distribution shifts, it mitigated accuracy decay by 34%.
For multivariate industrial telemetry, I modeled relationships across 30+ sensor channels with dynamic graphs and a two-layer Graph Attention Network. The anomaly engine used Peaks-Over-Threshold calibration and attained a 0.91 F1-score on industrial benchmarks.
I also quantized model representations to INT8 for sub-10ms inference per timestamp on standard CPUs, without relying on high-end GPUs. I’m pursuing a B.Tech in Computer Science and Engineering, specializing in Artificial Intelligence and Machine Learning, at VIT-AP University.
