Skip to content
View ramaseshanms's full-sized avatar

Block or report ramaseshanms

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
ramaseshanms/README.md

Ramaseshan Subramanian

LLM inference engineer. I get models running 3× faster on hardware you already own.

ramaseshanms.github.io  ·  LinkedIn  ·  m.s.ramaseshan@gmail.com


What I've shipped

Result Context
300% throughput on NVIDIA Jetson Orin/Xavier CUDA kernel optimization — directly enabled a Flash LIDAR partner to close a Rivian production contract
3.3× model compression, −1.06% perplexity Q4_HQQ quantization natively into llama.cpp · 151 tok/s on CPU inference
~3× visual SLAM speedup on constrained edge hardware ORB-SLAM3 memory architecture redesign for TI TDA4VM DSP via ROS2
Production AV perception in Autoware Universe BEVDet/BEVFormer real-time multi-sensor fusion with deterministic latency

Projects

Project What it does The number that matters
Viking Vulkan compute LLM runtime for Android 9 quantization shaders · 37 security fixes shipped
llama.cpp Q4_HQQ Native HQQ quantization with AVX2/NEON SIMD kernels 151 tok/s · 5.99 GiB → 1.79 GiB · 45% faster GGUF conversion
ORB-SLAM3 Edge Visual SLAM ported to DSP hardware, memory pipeline redesigned ~3× speedup on TI TDA4VM
EnterpriseGradeRAG Production RAG system with role-based access control 167 unit tests · multi-tenant
BEV Perception BEVDet/BEVFormer integrated into Autoware AV stack Shipped to production autonomous vehicle hardware
DeepCache Adaptive KV cache tiering between GPU VRAM and host RAM In progress

Stack

Kernel / Hardware: CUDA · Vulkan Compute · GLSL/SPIR-V · AVX2/AVX-512 · ARM NEON · OpenCL
Languages: C++20 · Python · Bash · ARM Assembly · Kotlin
Inference: llama.cpp · vLLM · ONNX Runtime · PyTorch · TensorFlow Lite
Platforms: AMD EPYC (NUMA, multi-socket) · NVIDIA Jetson Orin/Xavier · TI TDA4VM · Qualcomm Adreno
Tools: CMake · ROS2 · Autoware Universe · Docker · Perfetto · ZenDNN


Writing

I publish technical deep-dives on LLM quantization, Vulkan optimization, edge inference, and ARM porting — ramaseshanms.github.io.


What's next

The intersection of fast inference and agentic AI is where I'm focused. Multi-step agents are only as capable as their underlying compute is fast and cheap — and that's a kernel-level problem. Actively learning, building, and looking for problems in this space.


Open to

Founding engineer and AI infrastructure roles.

Best fit: teams where inference speed is a real constraint — edge AI, on-device LLMs, autonomous systems, agentic systems, or serious LLM serving infrastructure. I work best when the hardware problem is the product problem.

m.s.ramaseshan@gmail.com

Pinned Loading

  1. transformers101 transformers101 Public

    HTML

  2. autoware autoware Public

    Forked from autowarefoundation/autoware

    Autoware - the world's leading open-source software project for autonomous driving

    Shell

  3. llama_cpp_benchmarks llama_cpp_benchmarks Public

    C++

  4. llama.cpp llama.cpp Public

    Forked from ggml-org/llama.cpp

    LLM inference in C/C++

    C++