I've been building software since 2008. Today I design the microservice and event-driven architecture of high-traffic systems across the PHP, Go, Python and JavaScript ecosystems — focused on the problem, not the language.
What I learn from that work ends up across a few channels, each with a clear role.
| Channel | What you'll find | Language |
|---|---|---|
| muhammetsafak.com.tr | Applied development journal — what I built, how I built it, what broke | TR - EN |
| sade.dev | Architecture manifesto — production backend & systems decisions, the trade-offs behind them | TR - EN |
| YouTube | What text can't carry — process, side-by-side comparisons, walkthroughs | TR |
The line between the two sites is intentional: the journal is applied — granular, project-by-project, in the first person. sade.dev is the principled side — timeless, impersonal, the architecture I'd defend in a design review.
Architecting the event-driven, microservice backbone of high-traffic systems at Bulutklinik — designing for failure, scale, and the long run. Writing on muhammetsafak.com.tr since 2014, one month at a time — distilling production decisions into something worth reading.
- CommitBrief (github) — Local LLM-powered code review CLI for git diffs. Review staged changes, commits, branches, or PRs in your terminal — provider-agnostic (Anthropic, OpenAI, Gemini, Ollama, or your local Claude/Gemini/Codex CLI). Written in Go, GPL-3.0.
- BabelQueue (github) — Language-agnostic message queue standard: one strict JSON envelope that Laravel, Symfony, Go, Python, Java, .NET and Node.js all speak — over the broker you already run (Redis, RabbitMQ). Drops PHP's
serialize()lock-in, adds URN routing and built-intrace_idfor cross-service tracing. No sidecar or proxy. SDKs for six languages, MIT-licensed. - QueryProxy (github) — Self-hosted access control layer between developers and production databases: nobody gets credentials, everybody gets a query. AST-based SQL guards reject unprotected
UPDATE/DELETEand inject row limits, DBAs approve from the web UI or Slack/Teams, results come back masked-at-write with an immutable audit trail. PostgreSQL, MySQL, MariaDB, SQL Server and SQLite. A single Laravel monolith, no external services required, AGPL-3.0. - Contextator (github) — Turns a pile of documentation into a private MCP endpoint your agent can actually search. Point a project at local directories, git repos, uploads or Notion; it does markdown-aware chunking with heading breadcrumbs and incremental re-indexing by file hash. Embeddings run locally on CPU by default — OpenAI is optional, not assumed. TypeScript, PostgreSQL + pgvector, one Docker container, AGPL-3.0.
- RAGmux (github) — Self-hosted AI gateway: one OpenAI-compatible
/v1/chat/completionsendpoint in front of OpenAI, Anthropic, Gemini, DeepSeek, Ollama or anything else that speaks the protocol — with RAG, per-project rate limits, token budgets and audit logging built in. Two containers, no Redis, no separate vector database: users, documents, vectors and metrics all live in one PostgreSQL. Written in Go, AGPL-3.0. - Regnarium — A 4X strategy game that respects your evening: a full conquest on a 480-hex map in 30–45 minutes. Diceless combat that shows the damage before you commit, three AI personalities that don't cheat, seed-based maps you can share and replay. Runs entirely on-device — no servers, no accounts, no telemetry — in the browser today, iOS and Android next.
Most of what I open-source is built outside of work — because I enjoy it. If something here helps you, a ⭐ is always appreciated.
For everything else I've built, see the Portfolio page. I also gather what I build under a few umbrellas, each with its own remit:
- Parva Machina (github) — A one-person game studio, no publisher and no funding. Browser games — strategy, puzzle, simulation — that stay playable in airplane mode: no accounts, no telemetry, no hidden dice. Regnarium above is one of them.
- Tignex (github) — The engineering organization behind the developer tooling: self-hostable, forkable, built for systems already in production. Every project keeps its own name and repo; the decisions and trade-offs behind them are written down in public.
- Tunedness (github) — Applied AI research and engineering. Most AI projects stall in the last 10% — fine-tuning, retrieval, agents, evals, cost — and that's where the work starts. Everything ships behind eval gates, and the client keeps the weights, prompts and infrastructure.
Languages
Frameworks & tools
Data & infrastructure
The fastest way to reach me is info@muhammetsafak.com.tr.



