AI systems builder. I design, test and run multi-agent systems that keep their evidence, review each other's work and leave the final call to a person.
In the platform's five-week production phase I designed and ran an AI-assisted rebuild of its multilingual content: about 500 pages, 300 articles and 3,300+ FAQ entries. About 530 structured claims sit apart from the page text, each linked to its sources. Claude, GPT, Gemini and open-weight models reviewed each other's drafts, and clinicians approved clinical content before it was published.
For a pre-launch ride marketplace I ran about 40 agent lanes over two nights. They shared scarce Android test devices through atomic locks, and every finding needed a second reproduction before it counted. Night one gave 109 findings, 21 of them high severity, which became 70 tickets. A later 53-hour live-streamed soak test drove a booked ride with synthetic GPS on containerised Android and ended with 0 dropped frames.
I set up and benchmarked a CPU-only 35B mixture-of-experts model with llama.cpp on a 16-core, 96 GB ARM VM, then built a 159-task evaluation, 40% of it in Turkish, with pass thresholds fixed before the run. 140 of 151 tasks passed. An adversarial audit found 4 false passes, so I tightened the grader.
A method for parallel AI-agent browser QA with seven evidence rules. It ships with a demo shop that has eight seeded bugs, so you can score the agents against a known answer.
A Claude Code statusline with no dependencies. It shows cost, context use, rate limits, the git branch, MCP servers and system health.
One keypress copies Claude Code's answer to your clipboard as clean markdown instead of terminal drawing. Works over SSH through OSC 52.
A plain notepad in the browser for quick drafts and temporary notes. Vanilla HTML, CSS and JavaScript, with no login and no tracking.
Three browser extensions built on one promise: one click and the image is on your clipboard. No editor, no save dialog, no download, no account, no server. Each sibling captures from a different source, and they share an icon language so you can tell them apart at a glance. Together they have over 400 users and a 5.0 average from 32 ratings on the Chrome Web Store.
Click the icon and whatever fills the tab right now becomes a PNG on your clipboard. The one that set the family's rule. It also has a twin: the same extension ported to Firefox.
One click scrolls the whole page, stitches every section into a single tall PNG and copies it. For long articles, entire conversations and bug reports that never fit on one screen.
Hover any HTML5 video and the current frame is copied at the video's own resolution — only the video, never the page around it.
A terminal you can return to. Connect to your servers, work across tabs, and reconnect to the same remote session with server-side tmux.
I build software with clarity as the first principle. Code should communicate intent. Interfaces should feel intentional, not accidentally complex. Users should understand what's possible without reading a manual.