At National Institute of Technology (NIT), Warangal, I led research and development of ACAF-Net, a dual-stream RGB-Thermal fusion network for real-time edge surveillance. I designed its CSTF block and SAMG mechanism for dynamic cross-modal attention.
On the FLIR ADAS v2 dataset, ACAF-Net achieved a 48.1% relative improvement in mAP@0.5 and increased primary vehicle AP by 11.65%. I optimized the model with quantization-aware training and structured sparsity, reducing FLOPs by 51.5% and reaching 53 FPS INT8 inference.
I also developed a self-reflective multimodal RAG system that reasons over documents with text, tables, and charts. Its retrieval and fallback mechanisms reduced hallucinations by over 40%, with the system deployed through a FastAPI backend.
My other projects include an autonomous multi-agent LLM workflow and a video stylization engine using diffusion models. I hold an M.Sc. in Artificial Intelligence from Central University of South Bihar and work with Python, PyTorch, and tools including LlamaIndex and Docker.
