Ashwin Balasubramanian

Architecting a multi-tenant SaaS health platform with provider-agnostic AI routing, enterprise RBAC, Postgres RLS, and Langfuse observability. Engineering Fitbit OAuth 2.0 PKCE integration with encrypted token storage and resumable 90-day Inngest backfill workflows, alongside cache-first food search and privacy-focused data export/deletion flows.

Trained a unified deep learning pipeline across CheXpert, ChestX-ray14, MIMIC, VinDr-CXR, and ChestX-Det to classify, localize, and segment chest X-rays using a student-teacher contrastive learning architecture with cyclic pretraining to mitigate catastrophic forgetting. Applied ConvNeXt for classification, Faster R-CNN-ResNet for localization, and UNet-Swin for segmentation, achieving AUC 0.7348 on ChestX-ray14, mAP 0.115 on VinDr-CXR, and Dice 0.4711 on ChestX-Det using 2x A100 CUDA GPUs. Developed as part of a joint medical imaging research initiative between Arizona State University and Mayo Clinic, contributing deep learning solutions to assist radiologists in early and accurate detection of pulmonary conditions.