Swaminathan Sankaran

End-to-end system that detects silent model degradation through statistical drift monitoring and triggers automated retraining — recovering performance without manual intervention.

• Seven-stage Airflow DAG: data extraction, drift detection, retraining, smoke testing, model promotion
• Full containerized stack — FastAPI, MLflow, Evidently, Prometheus, Grafana — via Docker Compose
• Automated retraining recovered +0.1178 ROC-AUC on recession-shifted data

Tri-modal contrastive learning system for drug discovery — fuses 2D molecular images, 3D atomic geometry, and chemical fingerprints to predict pairwise similarity.

• Pretrained three encoders (ResNet-18, SchNet, MLP) on 291k molecules using NT-Xent contrastive loss
• Achieved ~0.92 Pearson correlation on expert-annotated pairs, beating the Tanimoto baseline
• Late-fusion architecture with per-modality cosine similarities for interpretable decomposition

Tamper detection for volumetric CT scans that identifies localized manipulation at the patch level — a harder problem than whole-scan forgery detection.

• 3D convolutional compressor converts 16-slice patches into 2D feature maps for ImageNet-pretrained ResNet-18
• Trained on 169 volumetric lung CT scans with 5-fold cross-validation
• Achieved 0.95 validation AUC, outperforming 2.5D, full 3D, and projection baselines