SELMA.CODES

selma kocabıyık — AI engineer

I buildAI systems thatactually work.

AI Engineer based in the Netherlands. BSc AI from VU Amsterdam. RAG pipelines, multi-agent systems, production LLM apps. Currently building vibeapp.

125K · 11 · VU Amsterdamscroll
Flagship · open source · MIT

Loop Engineering

Stop writing prompts. Start building agent loops. A Claude Code skill for self-running loops that keep going until a real, testable goal is met.

The gate is mechanical — a Stop hook that runs the actual test suite and costs zero tokens — so the agent can never approve its own half-finished work. Watch the koi: when it fails at VERIFY, it goes back to EXECUTE. That is the whole idea.

“The maker can’t be the judge.”
Claude CodeStop HookClosed LoopVerification Gate
runningattempt 1 · gate cost 0 tokens
move your cursor over the spacek = 6 · cosine
Inside the index

Every chunk is a point in here.

Retrieval is a geometry problem before it is a language problem. Real clusters, a query moving through them, and the six nearest neighbours lighting up as it passes.

This is what a retrieval bug looks like from the inside. When the wrong cluster lights up, no amount of prompt engineering downstream will save the answer.

embeddingsk-NNcosineclusteringsentence-transformers
Knowledge · agents

A vault that keeps writing itself.

An LLM agent reads new sources, integrates them into existing pages, updates the entity notes, flags its own contradictions and keeps the cross-references intact. I find the sources. It does the bookkeeping.

Hover a node — its neighbourhood lights up. That is a retrieval scope, not a decoration.

Obsidianpersistent wikientity pagescontradiction flags
30 notes · 48 linksliving index
Selected work — 2025 / 2026
Building in public — @selma.builds
01125Kyoutube views
0211builds filmed
034languages
041loop that runs itself