I’m Cristian Ramírez, PhD in Computer Science. My current R&D work focuses on performance engineering for generative AI: kernels, runtimes, benchmarking and agentic automation. I work with AI agents, vLLM, CUDA, C++ and Clang, with a background in research engineering at Universitat Politècnica de València and RISC-V vector kernels.
My engineering work connects AI kernels, runtimes, profiling and agentic performance optimization. My research background adds ARM NEON, RISC-V Vector, OpenMP and memory-aware numerical computing.
AI inference systems. AI agents and vLLM, with benchmarking, profiling and runtime optimization for the target hardware.
GPU software & compilers. Current tools include CUDA, C++ and Clang, connecting software implementation with execution behavior.
Computer architecture. Research experience in GEMM, ARM NEON, RISC-V Vector, OpenMP and memory hierarchies.
Projects
Engineering work grounded in implementation, architecture and published technical evidence.
@article{10620227,
author = {Ramírez, Cristian and Castelló, Adrián and Martínez, Héctor and Quintana-Ortí, Enrique S.},
journal = {IEEE Internet of Things Journal},
title = {Communication-Avoiding Fusion of GEMM-Based Convolutions for Deep Learning in the RISC-V GAP8 MCU},
year = {2024},
volume = {11},
number = {21},
pages = {35640-35653},
html = {https://ieeexplore.ieee.org/document/10620227},
keywords = {Convolutional neural networks;Tensors;Transforms;Internet of Things;Signal processing algorithms;Memory management;Matrix decomposition;Convolution;convolutional neural networks (CNNs);cost analysis;deep learning (DL);edge processors;high performance},
doi = {10.1109/JIOT.2024.3436937},
preview = {gemm_avoid_paper.png},
selected = {true},
bibtex_show = {true}
}
A research history of fast matrix multiplication, from bilinear identities and border rank to laser methods, 2026 claims, numerical stability and RISC-V GEMM engineering.
A practical guide to native Termux development and local CPU inference with llama.cpp on a Samsung S24+ with Exynos and 12 GB RAM, including GGUF verification, a local API and memory tuning.