Computer Programmer & IT Specialist
I build open-source developer tools for the AI era.
Small, sharp tools that measure how AI agents change a codebase, give agents better context with less noise, and catch problems before they ship. Written in TypeScript, Python, Go and Rust.
Featured work
Three tools that answer questions every team using AI agents is asking: how much of the code did AI write, does it last, and do agent pull requests get merged. Each runs on any GitHub repository and prints real numbers.
Does AI-written code get rewritten faster? Replays a repository's git history line by line and measures two-week churn, copy/paste and moved code, split between AI-assisted and human commits. Exact replay, checked against the real files. Single binary, no dependencies.
$ aichurn denoland/deno churn 7.1% of new lines rewritten or deleted within 14 days new lines churn AI-assisted 71,833 10% Human 581,811 6.7% ▲ AI-assisted lines churned 1.5× as often as human lines
How do AI agents' pull requests fare in your repo? Shows what share of recent PRs Copilot, Claude Code, Codex, Cursor, Devin, Jules and 8 more agents opened, how often those get merged and how fast, next to human PRs. Works on any GitHub repo; GitHub Action, SVG card and badge.
$ agentprs n8n-io/n8n 36% of the last 1,000 PRs were opened by AI agents PRs merge rate time to merge Claude Code 364 89% 23.9h Humans 452 88% 23.2h = Agent PRs get merged about as often as human PRs
How much of this repo did AI write? git blame for the agent era: the share of the code at HEAD written in commits signed by Claude Code, Cursor, Copilot, Codex, Aider, Devin and 9 more agents. Works on any public GitHub repo, writes an SVG card and badge, and can gate CI. Zero dependencies.
$ npx github:celaldrg/aiblame obra/superpowers 32% of the code at HEAD was written with AI 87 of 660 commits (13.2%) were AI-assisted Claude Code █████▏ 25.5% 9,083 lines Cursor █▍ 6.4% 2,289 lines
Developer tools
Single-purpose tools that fit into CI, cron and pre-commit. Every one ships with tests, a release and real example output.
Cuts the noise from command output before it reaches an AI agent. Passing tests, progress bars and install logs go; every error and summary stays, and the full output is one call away. MCP server and CLI. On its benchmark: 20,299 → 6,215 tokens (−69%).
$ shush go test -v ./... calc_test.go:26: Slug("Çay Saati") = "çay-saati", want "cay-saati" --- FAIL: TestSlugTurkish (0.00s) FAIL example.com/calc 0.002s [shush] exit 1 · 880 → 57 tokens (-94%) · hidden: 65 passing test lines
An MCP server with 17 Türkiye-specific tools for Claude, Cursor, VS Code and other clients: official TCMB exchange rates, live earthquakes, the holiday calendar and business-day math, IBAN/TCKN/VKN validation, provinces, prayer times and weather. No API keys.
celaldrg// ~/.cursor/mcp.json
{
"mcpServers": {
"turkiye": { "command": "npx", "args": ["-y", "github:celaldrg/turkiye-mcp"] }
}
}Linter for Agent Skills, the SKILL.md folders that Claude Code, Codex, Cursor and Copilot load on demand: name and description rules, strict frontmatter with line numbers, broken file links, context-size warnings. Zero dependencies; JSON and GitHub annotation output.
$ skillcheck examples/skills examples/broken ok pdf-forms (8 lines, ~73 tokens) FAIL Release_Notes (2 lines, ~16 tokens) SKILL.md:2 error name-format name 'Release_Notes' must be lowercase SKILL.md:7 error broken-link 'templates/notes.md' does not exist 3 skill(s), 3 error(s), 2 warning(s)
Linter for llms.txt, the file that points AI agents at your docs: structure, link format, relative and dead URLs, size of the Optional section. Works on a file, URL or whole domain; JSON and CI modes.
$ llmslint testdata/bad.txt # lines trimmed bad.txt:2: error [h1-first] the H1 title must be the first thing in the file bad.txt:2: warning [summary] no blockquote summary ("> ...") after the title bad.txt:6: warning [absolute-url] /install.html is relative bad.txt:7: warning [duplicate-url] /install.html is already listed on line 6
Readable diffs for lockfiles. Turns a 3,000-line package-lock.json change into what a reviewer needs: upgraded, downgraded, added and removed packages, major bumps first. Ten formats (npm, pnpm, yarn, Cargo, Go, Poetry, uv, Pipenv, Composer, Bundler), Markdown for PR comments, zero dependencies.
$ lockdiff 14.0.0 14.1.1 # ripgrep, rows trimmed Cargo.lock (cargo) 41 upgraded · 11 added · 9 removed · 2 major ↑ upgraded grep 0.2.13 → 0.3.2 major ↑ upgraded grep-printer 0.1.7 → 0.2.2 major ↑ upgraded aho-corasick 1.1.2 → 1.1.3 patch + added windows_* (8) 0.52.6
Know a CSV before you load it: type, missing values, distinct count, range, mean and top values for every column in one streaming pass. Detects delimiters, gzip and decimal commas, flags columns that will break an import. Parallel; about a second per million rows.
$ csvglance airport-codes.csv # columns trimmed airport-codes.csv · 86,286 rows · 13 columns · comma-separated · 8.4 MB · 87ms # COLUMN TYPE FILLED DISTINCT RANGE TOP VALUES 2 type string 100% 7 len 6 … 14 small_airport (42.8k), heliport (23.3k) 4 elevation_ft int 82.6% 6,485 -1,266 … 17,372 10 (597), 20 (469) 6 iso_country string 100% 247 len 2 US (32.7k), BR (8,113)
How old are your TODOs? Finds TODO, FIXME and HACK comments and uses git blame to show who wrote each one and how long ago, oldest first. JSON output and a CI gate for stale ones. Real output on ripgrep:
$ todoage --limit 3 # columns/lines trimmed to fit AGE AUTHOR LOCATION TAG TEXT 8y 6mo Andrew Gallant crates/searcher/src/sink.rs:607 TODO In theory, it should be possible to amortize 8y 2mo Andrew Gallant tests/regression.rs:179 TODO (burntsushi) Darwin doesn't like this test 8y 2mo Andrew Gallant tests/regression.rs:192 TODO (burntsushi) Figure out how to make this test work
Checks TLS certificates for one host or hundreds at once: expiry, hostname, trust chain and TLS version. Table or JSON output and exit codes for cron and CI. One binary, standard library only.
$ certpeek github.com expired.badssl.com wrong.host.badssl.com # columns/lines trimmed to fit STATUS TARGET EXPIRES DAYS NOTES CRITICAL expired.badssl.com 2015-04-12 -4215 expired OK github.com 2026-11-29 34 CRITICAL wrong.host.badssl.com 2026-12-28 63 hostname mismatch
A linter for .env files: syntax errors, duplicates, keys missing from .env.example, and real credentials committed to example files by mistake. Zero dependencies, with a pre-commit hook and GitHub Actions annotations.
$ envguard # columns/lines trimmed to fit error .env:4 unterminated-quote value of STRIPE_SECRET_KEY opens a ' quote that never closes error .env.example:6 secret-in-example AWS_ACCESS_KEY_ID in .env.example looks like a real credential (AWS access key) warning .env:6 duplicate-key SENTRY_DSN is also defined on line 5; the last one wins envguard: .env and .env.example: 3 errors, 4 warnings
Turkish text done right in Python: correct İ/ı casing, URL slugs, Turkish alphabetical sorting, numbers and lira amounts in words, 1.234,50 ₺ formatting. Zero dependencies, CLI included.
>>> import turkcekit as tk >>> tk.upper("istanbul ılık") 'İSTANBUL ILIK' >>> tk.slugify("Ağaç Ev & Bahçe Fiyatları 2026") 'agac-ev-bahce-fiyatlari-2026' >>> tk.sorted_tr(["zeytin", "çay", "ırmak", "irmik", "ördek", "ozan"]) ['çay', 'ırmak', 'irmik', 'ozan', 'ördek', 'zeytin']
Resources for developers in Türkiye: open-source projects, Turkish NLP, APIs and open data, communities, learning, podcasts, events and jobs. Links are checked automatically.
Playable in the browser
Small games and tools that run in any browser, phone included. Vanilla JavaScript, no build step, logic covered by node --test unit tests.
Full rules and an alpha-beta engine with three levels. PGN export.
Physics slingshot game on Matter.js, five levels of crumbling castles.
Puzzles with a guaranteed unique solution, notes, hints, three difficulties.
Sliding tiles with animations, multi-step undo and saved games.
Own expression parser, no eval: functions, DEG/RAD, history.
Safe first click, flags, chording, long-press to flag on mobile.
One-tap flying game with original art and a saved high score.
Brick breaker with three levels and power-ups such as multi-ball.
Four speeds, walls or wrap mode, swipe controls and a D-pad.
Unbeatable minimax AI, checked against every possible game.
Rules checked against perft counts, a three-level engine with a perfect endgame solve.
The traditional Turkish mancala game with its real rules and a three-level opponent.
Negamax with alpha-beta and a transposition table, running in a Web Worker.
Card-matching game: three board sizes, solo or two players on one device, 18 original SVG symbols and 3D flips.
Perfect mazes from three generators (backtracker, Prim, Kruskal with union-find), five levels and a BFS solution path.
Sliding puzzle from 3×3 to 5×5: every shuffle is solvable, IDA* hints on 3×3 and a picture mode.
Turkish backgammon against the computer (two levels) or a friend: bar entry, hitting, doubles, bearing off and mars.
Turkish draughts with flying kings and forced maximum captures, and an alpha-beta opponent in a Web Worker.
Hangman with English and Turkish word lists that knows ı from i, an animated SVG drawing, hints and streaks.
Nine men's morris with placing, moving and flying phases, a three-level computer opponent and drag, tap or keys.
Five in a row on a wooden 15×15 board: freestyle or exactly-five rule, alpha-beta AI with three levels in a Web Worker.
The classic Turkish card game against the computer: piştis, full scoring with a round summary, matches to 51, 101 or 151.
Vertical arcade shooter: ten scripted waves, a three-phase boss, a blaster plus three weapons that level up.
20 categories in English and Turkish, three sizes up to 15×15 in all 8 directions, every word hidden exactly once; drag, tap or keys.
Paint-by-numbers logic puzzles: 20 hand-made pictures up to 15×15 and endless random ones, each with exactly one solution and no guessing.
Crack the secret number, or let a minimax solver crack yours (any 4-digit one in at most 7 guesses); 3 to 5 digits, notes and answer checking.
Sea battle against the computer: drag five ships onto your sea, three AI levels, optional salvo and touching-ship rules, saved battles.
Klondike with Draw 1 or 3, numbered deals, a solver that can deal only winnable games, unlimited undo, hints and auto-complete.
3×3 to 7×7 boards against a friend or a computer whose Hard level counts chains, plays the long-chain rule and double-deals.
The Turkish trick-taking game with bidding, against three computer players that bid, pick trumps and remember the cards played.
The Turkish tile game against three computer players that read the discards, with a draggable two-row rack, seven pairs and false jokers.
Neon air hockey with swept collisions, a computer opponent on three levels that predicts wall bounces, and two players on one screen.
Go on 9x9, 13x13 and 19x19 with superko, handicap, area scoring and a Monte Carlo tree search computer player in a Web Worker.
Ten sorting algorithms step by step, with live counters, narration and complexity notes.
A*, Dijkstra, BFS, greedy search and DFS on a grid you draw, plus two maze generators.
Conway's Life on a canvas you draw on: zoom and pan, B/S rule presets such as HighLife and Day & Night, eight classic patterns.
Typing speed test with English and Turkish word lists: WPM, accuracy, consistency and a per-second chart.
Pixel sprites with mouse, pen or touch: fill, line, rectangle, mirror drawing, undo and PNG export up to 16×.
Focus sessions with short and long breaks, drift-free timing, a task list with estimates and a 7-day chart.
An 8-track, 16-step drum machine with synthesized drums, swing, patterns A–D, a song chain and share links.
A piano for touch and keyboard: four synth sounds, sustain, a recorder that keeps 8 takes and a Learn mode.
A private Kanban board: drag and drop for mouse, touch and keyboard, WIP limits, filters and undo/redo.
Harmonious palettes in OKLCH with locks, WCAG contrast checks, 50–950 shade scales and CSS, JSON or Tailwind export.
Live-preview Markdown with its own parser: synced scrolling, code highlighting, a table of contents and .md/.html export.
QR codes encoded from scratch with Reed–Solomon error correction: links, Wi-Fi, e-mail, SMS and vCards, PNG or SVG.
136 units in 14 categories with exact definitions; type “5 ft 3 in to cm”, see every unit at once, keep favorites.
Unbiased passwords and English or Turkish passphrases, an honest strength meter with crack times, and a checker for your own password.
Exact line and column errors, 20-digit numbers kept as written, a relaxed mode that repairs, a fast tree with JSONPath queries, and a diff.
Equal installments or equal principal with tax, fees, grace months and extra payments; schedule, chart, two offers and an affordability solver.
Live highlighting with colored named groups, a plain-language explanation of every token, replace and split previews, shareable links.
Its own parser, up to six functions with sliders, symbolic derivatives, roots, extrema, intersections and shaded integrals.
Mandelbrot, Julia, Burning Ship, Tricorn and Multibrot, rendered progressively on Web Workers, with a live Julia picker.
Line, word and character diffs with a linear-space Myers algorithm, a side-by-side view, unified patches and a patch applier.
A small spreadsheet with its own formula parser (no eval), 25 functions, a dependency graph, undo/redo and CSV import and export.
Draw any closed shape and watch a chain of rotating circles trace it, computed with an FFT; choose the number of terms and export SVG.
About
I'm a computer programmer and IT specialist from Türkiye. I like tools that do one job well: a single binary or a zero-dependency package, real example output in the README, and tests on every push.
Most of my recent work is about AI coding agents: measuring what they change in a codebase, and giving them cleaner context through MCP servers and linters.
Education & experience
Isparta University of Applied Sciences
Turkish Aerospace
Issues and pull requests are welcome on every project. Issues labeled good first issue are a good place to start. For anything else, reach out on LinkedIn.