SREE
DHARSHAN

Introduces a mean memory field that lets large populations of agents coordinate without pairwise communication. Against MFAC and MFQ baselines, memory-augmentation shows statistically significant gains (Cohen's d up to 4.9, p < 0.01, paired t-tests across 5–10 seeds), with the clearest population-size effect at N = 2 (+26.2%, d = 3.38) shrinking to non-significant by N = 8–16, and overhead scaling sub-linearly at O(N^0.33) up to N = 200. A conjectured O(1/√N) mean-field error bound is empirically consistent with the data (R² = 0.997). Currently under review; the open question is whether the memory field should be learned end-to-end or kept structured.

A novel event-triggered mechanism for decentralized multi-agent PSO-DMPC coordination that fires updates only when a local trigger condition is met. Cut inter-agent communication overhead by 92.5% and computational cost by 84%, with zero collisions across 10,000+ simulation timesteps in SUMO — improving the MTTC safety metric by 90.5% and reducing speed variance by 81.9% over baseline methods, validated across multiple traffic densities and parameter-sensitivity scenarios.

A YOLOv7 + ResNet-50 + Vision Transformer pipeline with spatio-temporal attention modules, benchmarked on FaceForensics++ and FakeAVCeleb. Reached 92% accuracy and 0.958 AUC while sustaining 25 FPS real-time inference, evaluating generalization across multiple forgery datasets rather than optimizing for accuracy alone.