claude setup review
May 30, 2026a free second pair of eyes on your claude workflow
hi, i'm mahir isikli. i build ai products end-to-end and put them in front of customers. a forward-deployed / product engineer based in berlin.
right now i'm building arbio: nexus, an ai ops platform running on 200k+ reservations, plus voice agents answering calls for 800+ properties.
before that, ai/ml engineer at intapp(nasdaq), shipping agentic workflows for fortune-500 finance & legal.
i'm a claude code ambassador, deeply embedded in berlin's ai ecosystem. i speak at ai tinkerers and organize regular meetups for 100-250 people. earlier i was president of TEDxTUBerlin and START Berlin.
i write about what i'm building and learning, and do tiny claude setup reviews for people who want their ai workflow to feel less annoying.

3,928 contributions in the past year, work included.
a free second pair of eyes on your claude workflow
lars loves vans. so instead of buying him vans, he got the entire vans catalog on his own body, swipeable.
a tiny mac script for the annoying problem where apple contacts birthdays do not show up in google calendar.
a tiny public mcp server for searching yorck showtimes, rendering seat maps, and letting my private agent book tickets safely.
A tiny experiment in pulling LibreLinkUp CGM data into a personal assistant and pairing it with a food diary.
organizing the first cafe cursor in berlin
every text model on openrouter dressed someone for the berghain door — tap any look for the exact prompt
every lineup from my favourite berlin open-air, matched to each dj's top track. auto-updates from the newsletter.
an animated laguerre-voronoi toy, inspired by jason davies
i see a lot of cool stuff every week. here's what caught my eye recently.
The best-in-class HTTP Debugger for macOS. Capture HTTP/HTTPS from macOS, iOS, Android with a few clicks ✅
Build RL environments for any domain from real cases that Browser Use agents mine from the web
Fully automatic censorship removal for language models
Reading diffusion language models for decisions: same OpenJev model, one pass instead of 3.65, +5 points, calibrated, no training. Plus cheap task branches on a frozen encoder.
The open-source docs framework for humans and agents.