Physicist and scientific software developer. I build simulation tools and developer tooling, and I check them against ground truth instead of eyeballing the output.

• tarhan — physics-first materials simulator, validation-first: 1D drift-diffusion, fuel cells and electrochemistry, cross-checked against DEVSIM and TiberCAD. The validation catalog is the product: manufactured-solution studies that measure the discretization order rather than asserting it. Archived with a Zenodo DOI. Live demo and run archive: tarhan-web.
• blender-optics-simulator — Blender add-on to design, simulate, align and render optical setups: live beam tracing, kinematic-mount auto-alignment, polarization, interference, Gaussian beams. The physics engine is checked against analytical oracles, not by looking at the render. On PyPI as blender-optics-mcp, listed in the MCP Registry, archived with a Zenodo DOI. Drivable from the UI or headless by an AI agent.
• dhm-hybrid — digital holographic microscopy workstation: off-axis reconstruction, quantitative phase imaging, autofocus, and reference-free phase retrieval through a hybrid classical + CNN pipeline. The autofocus default is pinned by a nine-scene real-lab benchmark, not by preference. Apple Silicon GPU paths, a PySide6 desktop app, and an MCP server so an agent can drive the pipeline headless.
• off-axis-hologram-sim — off-axis DHM simulator in MATLAB/Octave: split-step propagation through the object, angular spectrum to the hologram plane, a real 12-bit camera model, then reconstruction from the quantized counts alone. Synthetic holograms with known ground truth, for developing reconstruction code away from the optical table.

• optiproof — like code review, but it optimizes code and proves the speedup: differential testing for behaviour, Mann–Whitney + bootstrap CI for the timing claim. Python/Rust/Go, Docker-isolated.
• claude-skills — six Claude Code skills I actually use: multi-model delegation and routing, a three-agent loop that verifies its own work against rendered images, publishing that scans git history for secrets before it will push, and a skill that logs defects in other skills so workarounds get replaced by fixes.
• claude-i-know-kungfu — writes code in languages the model doesn't know, without hallucinating. Every symbol is grounded in real docs or a Docker-verified compile.
• claude-procurement-skill — procurement-grade product research: requirements, fair-price discovery, TCO, risk, decision memo.
• claude-continue — survive Claude's usage limit with same-chat continuity.
• claude-adhd-productivity — externalized executive function: one 2-minute next action at a time.
• claude-pessi-mystic — a pre-mortem engine that enumerates every way your project can fail.
• davinci-resolve-mcp-free — MCP server for DaVinci Resolve Free, which has no Studio scripting API.

patreon.com/c/emircbngl — how each problem actually got solved, what broke on the way, and early access to the tools before they ship.

One rule across everything here: no claim without measurement. If something is unverified, it says so.