I work on AI deployment at SNCF Voyageurs, where the question is not whether a coding agent is impressive but whether thousands of people can be trusted with one. That is a product problem, a security problem and a design problem at the same time, and it is the one I spend my days on.
Everything below started as something I needed and could not buy. I ship the tools rather than write about them, and I ship them finished: error handling, licences, a real README, a page you can send to someone.
Paris · open to conversations about applied AI, developer tooling and design engineering.
| What it solves | Built with | |
|---|---|---|
| Corporate Launcher | Turns any public AI coding CLI — Claude Code, Codex, Gemini, Cursor, Aider — into a company's own internal, audit-grade launcher. One structured interview produces 15 cyber controls, 30+ audit rules, a compliance .docx and universal cost tracking. The answer to "we cannot let a public CLI near our codebase". |
Python |
| Annotate Kit | Coding agents are blind to the screen. Point at an element — on a web page or in a running iOS app — write what should change, and the agent receives the element's identity, its surroundings and a marked screenshot instead of "the blue button, no, the other one". | TypeScript · Swift |
| SpecDrive | Spec-driven development, enforced. An MCP server and Mac app that walks coding agents through specs, scenarios, proof-gated tasks and independent review — because prompts get forgotten and tools cannot. | TypeScript · Electron · MCP |
| gptimage | Image generation inside Claude Code billed to a ChatGPT subscription instead of an API key. MCP server, skill and CLI in one. | JavaScript · MCP |
| imessage-connector | Gives a coding agent read, search, send and react on macOS iMessage — behind a policy engine whose layers can only ever narrow, never widen. | TypeScript · MCP |
| fable-mindset | A model's working discipline — decomposition, verification, orchestration, autonomous loops — extracted into portable Claude Code skills anyone can run for free. | Shell · Claude Code skills |
| MacPulse | A macOS menu-bar monitor with a local-AI readiness gauge: whether this Mac can run the Ollama model you are about to ask it for, right now. One Swift file, zero dependencies. | Swift |
More in the repository list — TokenWatt measures Claude Code spend in toasts and washing-machine cycles, and Banana Slides generates decks from a ChatGPT login.
- The user's experience decides the argument. Not the framework, not what is convenient to implement. If a choice is worse for the person in front of the screen it loses, and it usually loses in the first five minutes rather than in review.
- Production or it does not count. Error paths, edge cases, licences, security, the README someone else has to read. A demo that cannot ship is a slide, not a product.
- Design is not the last 5%. Most of these tools exist because the good version of the idea was unusable, and the difference was interface, not capability.
- Constraints from the real job. Working inside a large, regulated company is what taught me that the interesting problem is rarely the model — it is authorisation, cost, auditability, and whether a non-technical colleague gets it on the first try.
Swift · SwiftUI · TypeScript · React · Next.js · Node · Python · Electron ·
MCP · Claude Code · Codex · AppleScript · GitHub Actions
- connectedmate.com — AI talks, podcasts and Apple apps
- cormeraie.fr — a bilingual editorial notebook
- @AlexCormeraie on X