I build the systems that make AI actually work.

A six-agent software team that turns a goal into shipped, reviewed code.

A planning roundtable argues out an approach, then hands off to parallel execution — each task on its own git branch and worktree, merged only after separate QA and security gates pass. Cloud reasoning runs through headless Claude and Cursor CLIs, with a local model as fallback and vector memory that carries lessons between runs.

The safe "hands" every agent acts through, so autonomous execution can't go off the rails.

An agent says what it wants; OpenClaw decides if it's allowed, runs it locked inside a project scope, and records it. Every call appends a line to a SHA-256 hash-chained audit log you can verify. Overwrites get backed up first, a dry-run mode changes nothing, and git and shell commands pass an allowlist that blocks anything that could escape scope.

A coding agent that runs entirely on my machine and writes tested code for free.

A ReAct loop that plans, writes, runs and fixes code until a deterministic finish gate passes, then returns a machine-readable result so an orchestrator never has to parse human text. It remembers lessons across runs, and all 186 of its tests run fully offline. AI Workforce uses it as its free local build engine.

Reads the world's feeds and sends me only the alerts that matter, with the reasoning attached.

It watches news, Federal Reserve and White House statements, SEC filings, and on-chain whale moves, then clusters the same story across outlets into one event, dedupes it across restarts, scores it, and lets a small model judge the close calls before it pings Telegram. Runs around the clock on FastAPI, Postgres and Redis with a watchdog that restarts it if it dies.

An agency in a box — work flows as cards through a live board, with AI helping along the way.

Every task is a card moving through stages; the moment someone moves one, everyone sees it and the next person's queue updates over Supabase Realtime. AI helps at four points — writing captions, drafting client replies, translating between Turkish and English, and processing media. The app and its worker run on a single free Oracle Cloud box.

A multilingual coaching platform with an AI coach that writes real training plans.

Coach discovery and matching, coach dashboards and applications, custom program PDFs, and an AI plan generator with its own evaluation harness so I can measure how good its output actually is. It ships in five languages, takes payments through Paddle, and reports errors through Sentry.

A media agency's site and content pipeline — and a real test bed for my own AI platform.

A dark, motion-driven marketing site backed by a Sanity CMS and a simple admin dashboard, so the agency edits everything without touching code. I built a lot of it by driving AI Workforce, my own multi-agent platform, against a real deadline.

A Windows app for building restaurant menus and signage that works with no internet.

A desktop studio that stores everything locally in SQLite, so a café can run it on one machine with no server behind it. Built on Electron and React with local accounts and bundled media playback.