Stuttgart, Germany · Available immediately
|
Software engineer specializing in hardware acceleration with CUDA and Deep Learning Optimization using Quantization, Kernel Fusion and Compilers.
01 · About
I write GPU kernels on the hardware, with the profiler open. My work spans dependency-free inference engines, hand-tuned CUDA libraries for point-cloud and image workloads, and production pipelines that combine PyTorch compilation with TensorRT to squeeze out every millisecond. Outside of GPU work, I've validated real-time embedded Linux behaviour on physical hardware and shipped ML pipelines for time-series generation in the automotive industry.
Hand-written kernels, zero-dependency engines, real-time constraints on embedded Linux (PREEMPT_RT).
CUDA & OpenCL from first principles — tiled matmul, convolution, KNN, spatial sorting, reduction.
CNNs, UNet, Transformers & graph networks deployed through ONNX Runtime, TensorRT, and TVM.
02 · Skills
03 · Projects
A dependency-free UNet segmentation engine with custom OpenCL kernels written directly against the hardware — no cuDNN, no vendor libraries. Reaches a 600× speedup over a single-threaded CPU baseline on real image data.
Hand-written CUDA kernels for tiled matrix multiplication, 2D/3D convolution, furthest point sampling, KNN, prefix sum, reduction, and histogram — with FP32/FP16 benchmarks against AVX-optimized CPU baselines.
Assisted on the open-source DDACS dataset (32,071 FEM simulations) published to DaRUS with Croissant-format metadata, and built a FastAPI/Streamlit self-service tool for the IEH simulation pipeline with authenticated job submission.
Migrated a Yocto-based automotive BSP from Kirkstone to Scarthgap within Azure DevOps CI/CD pipelines.Integrated the Linux PREEMPT_RT patch and additional real-time patches, benchmarking scheduling latency to validate deterministic behaviour before release.
04 · Master's Thesis
Universität Stuttgart · Feb 2026 – Jul 2026 · MT-3970
A DGCNN-based surrogate model with cross-attention, sinusoidal distance bias, and FiLM conditioning, designed to replace 1.5-hour FEM simulations of sheet-metal springback with a 15-millisecond prediction.
DGCNN backbone augmented with cross-attention, sinusoidal distance bias, and FiLM conditioning for point-cloud springback prediction.
Windowed Morton / Z-order spatial sorting to improve memory locality and throughput for K-nearest-neighbor search on point clouds.
Combined PyTorch compilation with NVIDIA TensorRT, benchmarked against exact-KNN and library baselines (PyKeOps, cuML).
4× latency reduction per sample, from 60 ms to 15 ms.
05 · Experience
Validated real-time behaviour of embedded Linux (PREEMPT_RT) on physical hardware under strict timing constraints using JTAG/OpenOCD-based ARMv8 profiling. Hands-on hardware bring-up (Yocto BSP migration) and CI/CD pipelines (Azure DevOps).
Assisted on the open-source DDACS dataset (32,071 FEM simulations) published to DaRUS. Built a FastAPI/Streamlit self-service web tool for the IEH research simulation pipeline.
Benchmarked LSTM, Transformer, 1D CNN, and Seq2Seq architectures for synthetic time-series generation. Designed a custom loss aligned to ISO 18571, improving correlation accuracy by 10% over generic MSE.
Automotive-industry engineering experience prior to relocating to Germany for graduate studies.
Education
Universität Stuttgart, Germany · Oct 2023 – Jul 2026
Anna University, India · Aug 2015 – Apr 2019
Certifications
06 · Contact
Open to C++/GPU engineering roles. Reach out directly — I reply quickly.