∇Projects & field notes

Projects

Software built and maintained alongside client work, with AI coding agents under independent review. Client mandates appear here only as industry and role; the details belong to the clients.

§Field notes

Notes

How the tools are built, what the engineering decisions were and which rules hold in production.

The first field notes are being written. Check back soon, or ask directly.

§Own projects

01

LLM Harness

Agentic software analysis10/2026 – present

Connects language models to authorised repositories, documentation and tickets for source-backed analysis and controlled engineering workflows.

  • From research and proposed changes to controlled file edits, with separate permissions for reading, writing, test execution and Git operations.
  • Structured Robot Framework execution and read-only database and process diagnostics for investigating test results; resumable tasks with saved progress, bounded retries and recorded failure and recovery events.
  • Verified through automated Python and JavaScript tests, synthetic integration scenarios and independent reviews.

Python · JavaScript · MCP · REST · Robot Framework · Git · SQL

02

mnemo

Memory engine for AI coding agents05/2026 – 07/2026

Tiered short-, mid- and long-term memory across agent sessions.

  • Autonomous nightly consolidation runs with auditable reports and enforcement gates.

Claude Code · Codex · agent skills · MCP · launchd · Python

03

Comic Workshop

Agentic production line for a 114-page book09/2026 – 10/2026

Claude agents and an image model produced a 114-page comic and a 19-page making-of, all the way to the print PDF.

  • One agent strand per location, up to 16 pages in parallel; one source for every fact in YAML and a deterministic prompt builder with no AI in it.
  • Quality assurance as a mix of deterministic and agentic checks. A reader who never saw the story reconstructed it from the prompts alone: 957 of 978 story points.
  • Judge rules from production: no judge approves a repair it triggered, a missing answer counts as a failed check, new judges are first tried on known errors.

Claude Code · Codex · Python · OpenCV · YAML · Real-ESRGAN

04

Generative video studio

Local, on Apple Silicon06/2026 – 08/2026

A fully local studio for AI-generated narrated videos: images, video, narration and editing on the machine, with automatic quality checks.

  • Designed and built largely by AI coding agents.

Python · Remotion · SDXL · FLUX · LoRA · Wan 2.2 · Qwen3-TTS · Whisper · ffmpeg · MLX

05

robot-companion

VS Code extension for Robot Framework01/2026 – present

Editor tooling that surfaces return values, enums and types for Robot Framework, with its own Python backend for the full workflow in VS Code.

  • Documentation preview and return explorer in the sidebar, hover information and type-matched autocomplete from the indexed Robot and Python documentation.
  • Language server with navigation, diagnostics, completion, references, rename and formatting; test discovery in the Test Explorer; a debugger for mixed Robot/Python code.
  • Built with AI coding agents; published on the VS Code Marketplace under the MIT licence, the standalone backend as a preview release.

JavaScript · VS Code Extension API · Python · LSP · DAP · debugpy · Robot Framework

06

runcomposer & Test Manager

Test automation framework08/2025 – present

An API-first platform for orchestrating large automated test suites, and a Test Manager that lets business users assemble and run tests in plain business language.

  • runcomposer: tag-based selection with a compile step that previews exactly which cases a filter matches, run configuration and CI handoff. In use in a live test-automation mandate, MIT licence.
  • Test Manager (prototype, since 09/2026): business users with no knowledge of test code find, visually assemble and run automated tests in German or English, executed with Robot Framework. Tested on generated repositories of up to 1319 tests.
  • Developed agentically: coding agents implement tasks in isolated working copies; an independent review agent checks evidence and test results before a task counts as done.

Python · FastAPI · pydantic · React · TypeScript · Robot Framework · pytest · Jenkins