Edge AI / Computer Vision: YOLOv11, ONNX, HEF conversion, Raspberry Pi 5 + Hailo AI HAT, own datasets Embedded: ESP32 / M5Stack, ev3dev, OpenWrt, compiling drivers and untangling toolchains Linux & Systems: Arch, Nobara (Hyprland), self-hosted servers, networking RF / Mesh: self-built Meshtastic and LoRa nodes Languages: Python, JavaScript, Bash; currently learning C through an ESP32 project
ev3dev_butter-robot Real-time object detection on embedded hardware Split architecture: a Raspberry Pi 5 with a Hailo AI HAT does the vision, a LEGO Mindstorms platform running ev3dev does the moving. I collected and annotated my own dataset, trained YOLOv11, and converted it through ONNX to HEF for the accelerator. Camera and accelerator drivers had to be compiled by hand, and a HAT firmware update broke the toolchain badly enough that the whole conversion pipeline needed rebuilding. Detection runs reliably; the grabbing mechanism never fully worked.
clawdbot-twinmind-split-kit Self-hosted AI agent An agent running on a Linux server I administer myself: external API integrations, automated research and writing tasks, split between a conversational agent and a background worker. Later ported to an M5Stack microcontroller with voice control via Deepgram (pocket-sized terminal for the agent). Along the way I found that one integrated API was not serving the models it claimed to.
gambleandy-solana Solana token with automated fee lottery A Token-2022 contract that automatically distributes accumulated transaction fees to a randomly drawn wallet. JavaScript and the Solana CLI, running on testnet.
Meshtastic / LoRa mesh Several radio nodes built, configured and deployed around my area for text communication without cell infrastructure. Learned about range behaviour, power budgets and how nodes fail over a long winter the hard way XD.
linux-wallpaperengine (C++) improvements to render effect compatibility: transparent framebuffer initialisation, texture handling in render passes, and shader compatibility transformations for HLSL-style outputs. My original pull request was large, so the maintainers asked me to split it. It shipped as five individually reviewable PRs: #567, #568, #569, #570, #571 all merged. (my fork)
AI-2027 Paper Review and Optimized Forecast: An AI-Generated Audit, Methodology Critique, and Revised Forecast (Zenodo, April 2026) doi.org/10.5281/zenodo.19419882 A prediction audit and methodology critique of the AI-2027 study with an independent revised forecast for 2026–2032. The interesting part was the engineering: the dossier was produced by an autonomous multi-agent pipeline I built for it.
Android internals (custom ROMs, bootloader and boot image work), reverse engineering unfamiliar systems, and self-hosting whatever I would otherwise pay for.
- Learning C properly, through an ESP32 project, so I arrive at university with something usable rather than syntax knowledge
- Documenting older projects that never got a README | Hate docs :-(
- Looking for a working student position (Werkstudent) in embedded software or edge AI around Berlin and Potsdam, starting October 2026