osmancakir osmancakir

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

Osman Cakir

AI engineer in Munich. I build multi-agent systems and the products around them, which means I also write the frontend, the schema, the tests and the deploy pipeline. Most of my work sits where language models meet real collections: news, books, museum archives, research metadata.

Before software I studied economics and art history. That is why my agents tend to be evaluated rather than demoed.

Selected work

umbruchai-news-desk · LangGraph, TypeScript, OpenAI

A multi-agent newsroom that runs in production, and the backend of umbruchai.com. Five journalist personas research the day's stories in parallel and pitch; a human editor selects; the pipeline then writes each story at three CEFR levels of German, repairs its own schema and encoding errors through a validator agent, illustrates it, narrates it into three audio tracks, and publishes to a CMS. Parallel fan-out with state reducers, two human-in-the-loop interrupts, structured output captured through Zod-typed tools. Open source, including all seven agent definitions.

umbruchai · React Router 7, Cloudflare Workers, Sanity

The publication the news desk writes into: umbruchai.com, a German AI-journalism site now carrying 87 pieces across seven sections, each readable at three CEFR levels with audio at every level, and a page per agent author. Also open source, so the newsroom and the paper it prints can be read end to end.

candidgarden_web · React Router 7, pgvector, three.js, Cloudflare Workers (LMU Munich, with Prof. Dr. Kohle)

An institute for art re-search, now open source. Deployed at candidgarden.com, it holds 54,497 artworks from the ARTigo corpus, processed in three languages through hybrid human-AI verification, each read at Erwin Panofsky's three levels of meaning. The readings stay labelled provisional and confidence reports observed agreement rather than truth. Search embeds the query with bge-m3 and ranks it against 1,024-dimensional vectors in Postgres with pgvector, so a conceptual question finds work that shares no keyword with it.

The Atlas puts 89,800 of those readings into one WebGL point cloud. Its UMAP layout is fitted offline and never moves, so a search lights up matches inside a fixed space instead of rearranging it, and two searches stay comparable; a custom binary format ships the geometry in 1.6 MB rather than 8 MB of JSON. Server-rendered on Cloudflare Workers over Hyperdrive to a private RDS, with Vitest, Playwright and both runtimes built in CI.

ai-political-leanings · LLM evaluation, Python

Six frontier models answered the Political Compass's 59 propositions in English, German and Turkish: 1,062 answers, no refusals, no missing cells, scored on both axes. Five of six sit left-libertarian in every language, and the prompt language moves the result systematically, with German producing the most libertarian score for all six models.

Library Universe · React Router, TypeScript, Postgres, Sanity, fly.io

A library for reading and learning: a book catalogue and vocabulary flashcards on a spaced-repetition schedule. 249 commits, 225 TypeScript modules, Vitest and Playwright suites, GitHub Actions deploy. Source is private; the reading extension that pairs with it is not.

Städel Museum · multimodal LLM evaluation (freelance, 2026)

An AI metadata pipeline for artwork collections built on several multimodal models, with an evaluation workflow behind it: weighted scoring categories, blind comparison of model responses, and bias control conditions, plus a research interface for curators to review the output. The point was to find out which model an art-historical institution should actually trust, not to assume one.

agent-skills · Agent Skills for Claude Code and Codex

Installable skills, each with its own resources: a Bavarian public-sector job search that maintains a deduplicated database and scores ads against a candidate profile, and a cue-card generator that turns speaker notes into a printable duplex PDF deck.

DataCite Metadata Generator · React Router, TypeScript

A guided form that turns DataCite's deeply conditional kernel-4 schema into valid XML, with round-trip editing of records that already exist. Constraint-aware validation, so the repository stops rejecting submissions for reasons nobody can see. Live.

This is my own rebuild of a problem I worked on at LMU Munich, where I contributed 12 commits to the University Library's DataCite generator, which is in production at the UB.

What I work with

AI LangChain, LangGraph, RAG, pgvector, embeddings and UMAP projection, model evaluation, prompt and output-schema design, Claude Code, Codex
Frontend React, React Router, TypeScript, TanStack Query, Tailwind, Radix UI, Next.js, three.js and WebGL, TensorFlow.js
Backend Node.js, Postgres, SQLite and LiteFS, Python, Sanity, AWS S3 and RDS
Ops Docker, fly.io, Cloudflare Workers, AWS EC2 and CloudFront, GitHub Actions, Linux
Testing Vitest, Playwright, MSW, Sentry

Also

Five years of enterprise access-management dashboards at evolutionID GmbH, including a biometric photo capture module made roughly 50x faster by moving face detection into the browser with TensorFlow.js. DAAD scholarship, MSc Economics at LMU Munich.

Reachable at osmancakir11@gmail.com.

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!

osmancakir osmancakir

Osman Cakir

Selected work

What I work with

Also

Pinned Loading

Uh oh!