Skip to content
View RizgarOzan's full-sized avatar
🕷️
with great power comes great responsibility
🕷️
with great power comes great responsibility
6 followers · 8 following

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Achievements

Achievement: Pair ExtraordinaireAchievement: Pull Sharkx2Achievement: QuickdrawAchievement: YOLO

Highlights

  • Pro

Block or report RizgarOzan

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
RizgarOzan/README.md
Rızgar Ozan, AI systems and game tools programmer. I build games, tools and AI systems.

rizgarozan.github.io · LinkedIn · itch.io · Hugging Face · rizgarozan7@gmail.com

I build games, tools and AI systems: retrieval evaluations that say which part of a RAG pipeline earns its cost, and game tools that score a level or catch a broken font before anyone ships it. Final-year Computer Education & Instructional Technology student at Hacettepe University, Ankara.

111 pull requests merged into 36 projects I don't own — mteb, kornia, unity-mcp and docling among them — with 27 more in review.

rizgarozan.github.io: a moonlit field where each sword standing in the ground is one of my projects; scroll down into the smithy for what I'm working on now.

Selected work

01 Turkish RAG Eval — 3 chunkers by 4 retrievers on a hand-labelled Turkish gold set. Best nDCG@10 per retriever: BM25 0.41, BM25 + 5-char prefix 0.51, hybrid RRF 0.61, dense (multilingual) 0.50, dense (Turkish model) 0.78.
02 Match3 Lab — C# match-3 rules core, readable level format, Unity editor tools and a bot simulator. The greedy bot's win rate falls from 100% on level 1 to 50% on level 6; the random bot from 70% to 3%. 1000 games per level. 03 TMP Glyph Audit — Unity package that finds every TextMeshPro text the fonts cannot draw, through TMP's real fallback chain. Editor window plus CI exit codes.
04 Bilim Dedektifi — short educational mystery game, team course project at Hacettepe, free on itch.io. 05 Code-Enigma — React 19 academic progress platform built with İlkhan Arda Akmaca; I built the dashboard, peer comparison markers and the filter and settings controls.

Open source

Fixes sent back to the retrieval and Unity tooling I use, refreshed daily by a workflow.

Pull request Project
🟣 merged fix(geometry): accept empty batches in spatial_softmax2d and spatial_expectation2d kornia/kornia · ★ 11.4k
🟣 merged fix: keep the benchmark in BenchmarkResults.join_revisions embeddings-benchmark/mteb · ★ 3.5k
🟣 merged fix(augmentation): make auto policies picklable kornia/kornia · ★ 11.4k
🟣 merged [fix] Warn about every include_positives option that discards the positives huggingface/sentence-transformers · ★ 19.2k
🟣 merged [fix] Keep leading dims when binary-quantizing a single embedding huggingface/sentence-transformers · ★ 19.2k

…and 106 more merged pull requests. Merges by project: mteb 19 · kornia 17 · unity-mcp 8 · docling 6 · browser-compat-data 5 · results 4 · LightRAG 4 · module-replacements 3 · telegram-mcp 3 · sentence-transformers 3, and 26 more projects.

Now · October 2026

  • Turkish RAG Eval 0.2.0 is on PyPI: 300 queries (58 hand-labelled, 242 LLM-drafted and double-labelled) across five models, with the dataset on Hugging Face. The MTEB task is in review (mteb#5615), and the drafted queries are open for human review as good first issues.
  • Graduating in June 2027; open to internships and new-grad roles in applied AI and game engineering.

Pinned Loading

  1. tmp-glyph-audit tmp-glyph-audit Public

    Find every TextMeshPro text your fonts cannot draw - scenes, prefabs and runtime text, through TMP's real fallback chain. Editor window + CI exit codes.

    ShaderLab 1

  2. turkce-humanizer turkce-humanizer Public

    Türkçe yapay zekâ metnini, ne dediğini değiştirmeden yazarın sesine çeviren ajan skill'i (Claude Code, Codex, Cursor). Humanizer'ın Türkçe karşılığı.

    Python 3

  3. match3-lab match3-lab Public

    Match-3 workbench: pure C# rules core, readable level format, bot simulator that measures difficulty, Unity editor tools + WebGL demo

    C# 1 2

  4. turkish-rag-eval turkish-rag-eval Public

    Which parts of a RAG pipeline earn their cost in Turkish? 12 retrieval configs measured on a hand-labelled Turkish gold set

    Python 1 1

  5. konsey konsey Public

    Claude Code karar konseyi: dört danışman ve bir hakim bir fikre BUILD, FIX FIRST ya da KILL der