Projects

Built an edge-deployable deepfake image detector using ResNet-50 transfer learning with quantization for deployment. Achieved 82% accuracy on the quantized model while tracking latency and size trade-offs to keep inference practical on constrained devices.

Implemented inference-time loss functions for severe color control, monitoring per-step ROI drift and reconstruction error to preserve target color fidelity while keeping the background unchanged. Achieved ΔE₀₀ under 5.

Built a Vision Transformer from scratch with patch embeddings, attaining 75% test accuracy with 12% improvement over CNN baseline. Reduced training time by 22% via gradient checkpointing and implemented LR warmup with cosine decay to converge 3× faster than standard schedulers.

Synthesized mandatory and voluntary reporting dimensions using fairness-constrained optimization to derive weightings penalizing asymmetrical disclosures. Collected over 68 transparency items, verified scores against Guidestar India rankings, and led a team of 3 in collaboration with CSIP at Ashoka University.