neuropulse
Live 3.8B-param forward pass in the browser with mech-interp ablation.
portfolio · barisgunaydin.com
41 open-source repos · 1,833 commits in last 90 days · WebGPU kernels, browser-native quantum chemistry, local-first AI infrastructure. Independent research.
synced 2026-06-24 · interactive view loads in <1 s with JS
Live 3.8B-param forward pass in the browser with mech-interp ablation.
WebGPU quantum many-body + chemistry simulator. Brute-force verified element-wise against PySCF, with WebGPU acceleration on the CCSD(T) bottleneck.
The editor you're using right now.
Federated SGD over Cloudflare Durable Objects. 50% Byzantine tolerance.
TR-MRV-Bench — open emissions benchmark beating EU CBAM by 85.9%.
Phi-3-mini running in a browser on 10 hand-written WGSL kernels. No TVM, no compiler, no API key — within striking distance of the autotuned baseline for ~40× less code.
Research umbrella for GPU kernel fusion — eliminating per-dispatch overhead across evolutionary computation and transformer inference. Two preprints.
Real WebGPU compute benchmarks running on visitors' hardware — kernel-fusion validated across hundreds of devices and seven GPU vendors.
Browser-native LoRA fine-tuning of BitNet b1.58 2B4T — first public demo of training a ternary 2B LLM in a web page.
Drop-in npm install replacement that sandboxes postinstall scripts, blocks network access, detects typosquats, and catches supply-chain attacks before they execute.
Local-first grammar + writing assistant for macOS. Native system-wide overlay across every app. Harper rules + on-device Gemma/BitNet via QVAC Fabric. 100% on-device, zero cloud.
Radiobiology Monte Carlo in a browser tab — WebGPU port of CNRS Geant4-DNA. Karamitros IRT chemistry + 21×21 B-DNA fiber-grid scoring, cross-checked against the canonical reference.
Multi-agent rights advisor for APJ migrant workers — five specialists read your contract in your language, visibly disagree, then converge. Databricks APJ hackathon entry.
Open scientific questions in plain words — a 3D lab where a researcher answers anything, draws it on a whiteboard, ties it to a real benchmark. Browser-only, bring your own LLM.
Privacy layer between any app and any third-party LLM — pseudonymizes emails / paths / URLs / IPs deterministically before the wire and reverses the mapping on the response.
Realtime LIF simulation of the FlyWire connectome + MANC spine + flybody body — a fly brain talking to a fly body, live in your browser via WebGPU + WASM.
Online context-healing loop for long-running LLM agents — tag-then-rebuild compression preserves cross-chunk coreference instead of the cliff-edge auto-compact.
Realtime in-browser Stable Diffusion with a fused U-Net pass — single-kernel attention + conv + groupnorm + tembed.
Local-first AI vault for Turkish attorneys — KVKK-compliant: müvekkil dosyaları cihazdan çıkmaz. Derivative of markview targeting Turkish solo/small-firm lawyers.
Multi-agent advisory panel for early-stage founders — four AI agents discuss your idea in a 3D meeting room. Local LLM, no cloud, no telemetry.
Local-first browser brain — SolidJS + WebGPU on Cloudflare. Persistent agent in the right rail; privacy layer chooses which model sees what before the message ships.
WebGPU vision training engine + runtime shader generator + gradient-free ES variant. Browser-native image classification on Malaria Cell Images. Companion to fused-lora — the vision-side substrate.
A tiny local HTTP + SSE server that exposes a markdown vault and a fleet of agent sessions to any browser tab, extension, or IDE on the same machine.
A 3D obsidian-style knowledge visualizer. Drag-drop markdown, PDFs, .docx, .xlsx — they orbit each other by wikilinks, on-device embeddings, and TF-IDF semantic similarity.
An embodied 3D presence layer for the web — face mesh, particle nebula, voice bridge. Drop into React + Three.js for oracle / avatar / guided-experience UIs.
Give Claude Desktop, Cursor, ChatGPT, or any MCP-aware app read-only access to your local files — without uploading anything. Electron + WebRTC + stdio MCP.
Fused WebGPU compute kernels for transformer inference — 66-458x faster
Browser-to-browser distributed evolution. WebGPU compute + WebRTC P2P. 4 demos, 113-line relay. No install, no CUDA.
Single-kernel fusion for autoregressive transformer decoding via WebGPU
ADAS sensor fusion benchmark — 11-stage fused wgpu-native vs multi-kernel PyTorch. 12-15x faster on same GPU.
Pushing fused WebGPU transformer kernels to max model size — int4, tiled FFN, Phi-3-mini 3.6B in Chrome
wgpu-native benchmarks — same WGSL shaders, no browser. For CUDA vs wgpu comparison.
Research umbrella for GPU kernel fusion — eliminating per-dispatch overhead in sequential GPU workloads, from RL to transformer inference.
A medieval village simulation — 15 souls, named buildings, deterministic worldgen. TypeScript port of swarm-engine/demos/town.html.
Teklifçi — on-device VLM for Turkish metal-fab quote extraction + machine-engineer collab
Zero-dependency replacements for popular npm packages. Copy. Paste. Own your code.
Browser-based protein sequence evolution via WebGPU. Honest about proxy-fitness scope (no wet-lab validation).
Aggregation strategies for distributed evolution — six political-system metaphors tested on crypto trading.
Privacy-preserving distributed AI via browser-based Evolutionary Strategies — measured Byzantine tolerance at 50% attacker share.
Directed-evolution engine for dual-function enzymes — PETase + metallothionein fusion for a Chiang Mai water filter.
Chiang Mai burning-season simulation — the project that triggered the kernel-fusion line.