What I have built, newest first. Most of the code is private, so each entry says what is public and what is not.
A desktop control room for a team of AI agents: Claude Code, Codex and Grok as command-line sessions, reachable from my phone, each optionally with its own Docker Linux desktop.
Over 1,300 automated tests. Every release ships with a rollback. No outside users yet.
Read the case study: Agent FloorAn MCP server that gives agents across my machines one keyword index over shared notes, with token auth and gitleaks redaction before indexing.
9 of 10 recall test questions answered in the top 5 at build. Fixes from an independent security review.
Read the case study: mesh-memoryHeadless Claude workers that share one cached prompt, with a model and effort level per task, hard spending caps and a cost ledger.
94% cache hit on a test relaunch. In an escape test, worker commands could not write outside their folder, read SSH keys or reach the network.
Research run by teams of Claude Code sub-agents: researchers and adversarial reviewers, with cited sources.
In a 13-agent email-gateway study, all six reviewers found the same flaw: no process boundary between the agent and the mail credentials. The design was parked.
A six-stage workflow that turns job photos into draft social posts, gated by a separate Claude judge with a rubric and hard-fail rules, then by human approval.
Dry-run tested on a made-up client. It never went live, and the client relationship ended.
An open-source screenshot tool for Linux: Wayland-first capture, OCR, an annotation editor, a clipboard daemon run as a systemd user service, an optional GNOME Shell extension and a distro-aware installer.
Core modules are tested headless with pytest.
See the code on GitHub: ClipShotAlso: florinstudio.uk, the site for my one-person studio, built in React and Vite. Its portfolio pieces are labelled as concept demos.