AI and HPC applications are increasingly reliant on multi-GPU collective communication and fused compute-and-communication patterns. AIRamp‑Accelerate is a C++/HIP framework for AMD ROCm that applies Policy‑Based Just‑In‑Time (PBJIT) dispatch to select and run the best communication/computation strategy at runtime. Concretely, AIRamp‑Accelerate implements optimized kernels for All‑to‑All, Allgather+GEMM, and GEMM+ReduceScatter, using RCCL for collectives and rocBLAS for matrix math, compiled with HIP (hipcc) on ROCm 6.x+.