ARYAN

96.51% accuracy, 0.998 macro AUC, and 0.954 macro F1 on 687 unseen patient images, split at the patient level and verified leakage-free. An EfficientNet-B3 brain MRI classifier that returns per-class confidence and a Grad-CAM heatmap.

Predicts the year-over-year change in a JoSAA cutoff rather than the raw rank, then ranks the colleges a rank can realistically reach. Admission probability is calibrated with isotonic regression on a held-out year and rounded to 5%.

100% detection at a 0.131% false-positive rate over 90 randomized attack runs, from four independent detectors: rules, statistics, Isolation Forest, and physics. An agentic SOC for a simulated smart grid that keeps false-data-injection attacks out of dispatch.

Turns a UI screenshot into semantic HTML or JSX with Tailwind, streamed into a live sandboxed preview as the model writes it. Built on Next.js 16 and the Vercel AI SDK with Gemini vision, plus a refinement loop for iterating without re-forging.