Vulkan · CUDA · Neural Rendering · Agents · Simulation
I build renderers, physics engines, and the pipelines that feed them — from hybrid deferred-plus-ray-traced Vulkan to CUDA path tracers with MIS and OIDN denoising, down to the capture and reconstruction infrastructure that turns photons into assets.
MSc Computer Science, Linköping University — Graphics & Visualization. Ships hardware RT pipelines (Vulkan KHR, CUDA), MIS path tracing with environment-map importance sampling and Owen-scrambled Sobol, three production denoiser backends (Intel OIDN / NVIDIA OptiX / NRD), differentiable rendering (nvdiffrast + Disney BRDF), and synchronized multi-camera capture systems. Currently AI algorithm intern at Freedo Technology. Open to graphics / engine / GPU-systems / AI infrastructure roles — available full-time.
Damaged Helmet · deferred PBR · full material stack
Lantern · emissive driving scene lighting
MetalRoughSpheres · BRDF calibration
Path-tracing pipeline · Intel OIDN denoising on a glossy scene
Helmet · 16 spp · raw path-traced
Helmet · 16 spp · OIDN denoised
C++20 · Vulkan KHR RT · GLSL · ~50K C++ + 15K GLSL across 121 shaders
Standalone Vulkan engine written from scratch. Hybrid renderer: a deferred G-buffer path and a hardware ray-tracing path share one scene representation, bindless resources, and acceleration-structure workflow. Architecture spans render (22K LOC), physics (11K LOC), and a gpu / Vulkan abstraction layer (8K LOC).
Deferred path: Cook-Torrance GGX BRDF, cascaded shadow maps, SSAO, screen-space reflections, subsurface scattering, TAA, bloom. Path-tracing path: Vulkan KHR ray-tracing with multiple-importance sampling (balance + power heuristic), environment-map importance sampling via precomputed marginal/conditional CDFs, Owen-scrambled Sobol QMC, variance-driven adaptive sampling, and three denoiser backends (Intel OIDN, NVIDIA OptiX, NVIDIA NRD).
Physics layer: GJK / EPA collision detection over a BVH. Top renders use Khronos glTF Sample Assets — a sanity-check that the engine reproduces the reference results graphics engineers actually verify against.
Gränsö Castle · aerial photogrammetric reconstruction
Method matrix · RC / NeRF / 3DGS · glass, stone, skull
C++ · C# · WinUI · Hardware Control · ~8K lines
Built CamMatrixCapture: a 12-camera synchronized capture system with staggered trigger firmware, Bluetooth-controlled turntable automation, and a state-machine driven capture workflow. Indoor object captures feed NeRF / 3DGS training; outdoor work includes drone reconstruction of Gränsö Castle.
Released a 209 GB public dataset focused on challenging materials — reflective, transparent, translucent surfaces that traditionally break neural rendering methods.
The benchmarking contribution: 3D Gaussian Splatting reaches ~11 dB PSNR above NeRF on reflective surfaces — a surprisingly large gap, with clear implications for production capture pipelines.
CUDA path tracer with BVH, Schlick-Fresnel refraction, and optional Intel OIDN denoising. 16 parameterized liquid presets (ocean · honey · lava · milk · blood · crystal …) each with their own IOR and depth-absorption. Paired with a Taichi-GPU Position-Based Dynamics cloth solver — stretch + bend constraints, material-specific stiffness — exporting per-frame OBJ meshes back to the CUDA renderer. 58 production stills shipped.
Repo ↗
Foundation for a differentiable rendering pipeline that will learn per-point BRDF from turntable captures, then relight glossy objects uniformly so Structure-from-Motion actually converges. Full GGX stack shipped — D (GGX / Beckmann / anisotropic), G (Smith / Schlick-GGX), F (Schlick / roughness-aware) — plus nvdiffrast backend, turntable camera / mesh / lighting layers. Neural decomposition + training loop next. Built on NVIDIA nvdiffrast.
Repo ↗
GPU-accelerated Smoothed Particle Hydrodynamics solver with Marching Cubes surface extraction and CUDA ↔ OpenGL interop for real-time visualization. Uniform-grid spatial hashing for neighbor search, cubic-spline density kernel with gradient / Laplacian variants for pressure and viscosity forces, plus surface-tension kernel. Bachelor thesis — the foundation my current fluid work builds on.
Repo ↗ Demo video ↗
Compute-shader playground for procedural material decay on a PBR sphere — noise-driven rust progression (0/25/50/75/100%), paint cracking and peel, and a dynamic puddle / weather system that wets and dries the ground in real time. Built around an OpenGL compute pipeline with full BRDF response so albedo and roughness shift together as the surface ages. Coursework for Linköping's TNM084 Procedural Methods.
Repo ↗
Reinforcement-learning agent for Chinese Chess — full Xiàngqí rules engine in C++/Qt, a Deep Q-Network with experience replay and a target network, and CUDA-accelerated training driven by self-play in the spirit of AlphaZero. Win-rate sweeps against random and self-play opponents, plus a playable GUI to face the agent. TNM114 group project at Linköping with Oskar Tengvall.
Repo ↗
Two global-illumination renderers written from scratch for TNCG15. Monte Carlo path tracer with a CUDA backend for per-pixel parallel ray generation and progressive accumulation — real-time-feasible against the CPU baseline. Photon-mapping pass scatters photons from light sources, then estimates radiance via KD-tree nearest-neighbour lookup for sharp caustics and indirect lighting. Comparative study of where each technique wins.
Repo ↗ Demo ↗
Pipeline for classifying skiing techniques from video — OpenPose extracts 2D body keypoints, a Python preprocessor stitches per-frame skeletons into spatial-temporal graphs, and an ST-GCN trains on those graphs to label actions like big-bend carving. Best model (epoch 39, lr 0.1, batch 32) hits 50% Top-1 / 90.9% Top-5. Co-authored, published in Intelligent Computer and Applications Vol. 12 No. 4, April 2022.
Paper ↗I care about infrastructure, not app surface area. The systems I find interesting are the ones other engineers build on top of: a renderer, a capture pipeline, a physics engine, a GPU scheduler. I profile before optimizing, I read the paper before reimplementing, and I'd rather ship one well-understood 50K-line engine than three shallow demos.