Founding Engineer, Independent Researcher
Cited work first — released, DOI-archived, and validated against an external reference.
| zero-tvm · zerotvm.com · DOI | Ten models on hand-written WGSL, from Phi-3-mini up to a 35B sparse mixture-of-experts that no compiler stack ships a build of. Benchmarked against WebLLM on identical weights in the same session, every number in BENCH.md. Peer weight-sharing over WebRTC and pipeline splitting across two tabs. |
| webgpu-dna · webgpudna.com · DOI | The CNRS/IN2P3 Geant4-DNA Monte Carlo toolkit in a browser. Electron track structure, Karamitros IRT radiolysis, SSB/DSB scoring on a B-DNA fibre grid — validated against a Geant4 11.4.1 ntuple across eight energies, with the failed experiments committed beside the passing ones. |
| webgpu-q · live · DOI | Quantum many-body and chemistry — statevector and MPS, plus HF/UHF, DFT, MP2, CCSD, CCSD(T) and EOM-CCSD, cross-checked against PySCF, ITensor and brute-force EOM. A six-level research ladder with fidelity pass bars and seeded RNG. |
| neuropulse · neuropulse.live · DOI | A full Phi-3-mini forward pass rendered 1:1 from live activations — 3.8B parameters, 11 WGSL kernels, 292 dispatches per token. Every glow is a real tensor read back off the GPU. |
| webgpu-kernel-fusion · kernelfusion.dev · DOI | The research line: two preprints on single-kernel fusion for GPU compute, measured across CUDA, WebGPU, JAX/XLA and Triton with a CI-gated equivalence check between implementations. |
| iz · DOI | Per-facility CO₂ benchmark for Turkish CBAM-scope industry. A closed-form physics baseline against the EU CBAM default, validated on EUTL-verified plants under leave-one-plant-out. |
| gpubench · gpubench.dev | WebGPU compute benchmarks on visitors' hardware. Six workloads across eight GPU vendors; the run table is published as a dataset. |
Shipped products, on their own release cadence.
| markview · markview.ai | Local-first Markdown editor. Web app plus desktop builds for macOS, Windows and Linux. |
| nib | Local-first grammar and writing assistant for macOS. Native overlay across every app, Harper rules plus a bundled on-device model, no account. |
| webgpu-fly · live | Realtime LIF simulation of the FlyWire fruit-fly connectome, the MANC spine and a MuJoCo body — 139,255 neurons, ~15M synaptic edges. |
Earlier and experimental — live, but each carries its own caveats in its README.
enter-the-painting lifts any image into a 3D Gaussian-splat cloud, fully client-side ·
draw-instant runs Stable Diffusion in the browser, with the all-our-kernels path still being extended to the U-Net ·
fused-lora fine-tunes BitNet b1.58 in a tab and emits ~4 MB .flora adapters ·
wgpu-adas-bench fuses an ADAS sensor-fusion pipeline into one wgpu-native dispatch (archived) ·
safenpm scans npm install for typosquats and risky postinstall scripts, pre-release ·
dimples is a golf game where you are the ball
- Email: hi [at] barisgunaydin.com
- Personal site: barisgunaydin.com
- HuggingFace: @abgunaydin — models, Spaces, and the WebGPU compute benchmark dataset
- X: @abgnydn_
- Location: Chiang Mai, Thailand (UTC+7)