MYLIFEGRAPH / DESIGN SYSTEMFlutter · Personal daily planning
01 / Apps & Tools
MyLifeGraph
A personal companion for planning, habits and focus. Daily check-ins meet a clear view of everyday life.
FlutterDartPython
More about this project
A mobile-first coaching app built with Flutter, FastAPI and Supabase. Guest mode makes it possible to explore without an account; planning and guidance are based on explicit user input.
Personal weekly meal plans, matching shopping lists and upcoming meals at a glance. Also known as V&M Fuel.
JavaScriptMeal Planning
More about this project
A nutrition project built around personal preferences, weekly planning and shopping lists. The original V&M Fuel project website remains available in the archive.
WORKSPACE◈ Portfolio◇ Playground+ New projectCodex · Claude · Grok
✦ Codex
Build a useful app.
✓Understand the project
✓Implement changes
◌Review the result
An idea becomes an app.↑
Demo with sample content
03 / Web
Codex UI
A dedicated interface for agentic development: projects, chats and multiple coding agents in one place.
TypeScriptWeb UI
More about this project
A personal web app for orchestrating coding agents, managing projects and chats, and voice input. The visual preview is an illustrative demo with sample content.
Private source
YAPP AI / VOICE TO TEXT
Speak it. Keep it.
00:12
TRANSCRIPTION✦
Capture a new idea.
Give the thought room to grow.
Windows + AndroidDemo with sample content
04 / AI & Research
Yapp AI
Speak instead of typing: native voice input for Windows and Android, with separate meeting recording and transcription.
JavaC#Speech-to-Text
More about this project
A native application in development, featuring a microphone overlay, dictation history and configurable speech providers. Device validation is still part of ongoing development.
Private source
ZERO-SHOT / CHEST X-RAY
Image meets language.
Image
Text prompts
Shared embedding space
Multiple findings ↗
Research · model and prompt comparison
05 / AI & Research
Zero-Shot Chest X-Ray
Which findings can vision-language models identify without task-specific training? Comparing BioMedCLIP and MedCLIP with different prompting strategies on chest X-rays.
PythonVision-Language ModelsZero-Shot Learning
More about this project
A research project on zero-shot multi-label classification with BioMedCLIP and MedCLIP across CheXpert, NIH ChestX-ray14 and PadChest. We compared positive and negative prompts, LLM-generated descriptions and composed strategies. Evaluation shows that the benefit of a prompt strategy strongly depends on the model and dataset.
Private source
3D VISION / GAUSSIAN SPLATTING
Many views. One scene.
Concept illustrationCameras → 3D Gaussians → new views
06 / AI & Research
3D Reconstruction with Gaussian Splatting
A team project reconstructing 3D scenes from multiple calibrated camera images. A neural model produces a 3D Gaussian representation that can be rendered from new viewpoints.
PythonPyTorch3D Gaussian Splatting
More about this project
For the project, we developed a pipeline that fuses image features from multiple cameras to predict a metric 3D representation of objects and their surroundings. The work covered model training, novel-view rendering and evaluation of reconstruction quality.
Private source
TÜBINGEN / SEARCH ENGINE
Your city. Searchable.
Original project interface
07 / AI & Research
Tübingen Search Engine
A search engine for websites about Tübingen: crawling, indexing and ranking in a collaborative university project.
PythonStreamlitInformation Retrieval
More about this project
Created in the Modern Search Engines course. A pipeline turns websites into a searchable index and presents results through a Streamlit interface.
SAC ↔ Agent vs. agentTraining → Benchmark → Competition
08 / AI & Research
Hockey RL Agents
Developing and training reinforcement learning agents to face other agents in a provided hockey environment. The project concluded with a competition between teams.
PythonReinforcement LearningSoft Actor-Critic
More about this project
My contribution to the team: a Soft Actor-Critic (SAC) agent, training against a pool of different opponents and benchmarks to compare models. The trained agent competed against other teams’ agents in the final competition.
Private source
It starts with a sketch.
Sketch → Image
09 / AI & Research
Pencil2Pixel
From a first sketch to an AI-generated image. A drawing canvas meets prompts, visual styles and creative experiments.
PythonJavaScriptGenerative AI
More about this project
Draw in the browser, add a prompt and generate images through a Python backend. The project includes brushes, an eraser, undo/redo and adjustable generation parameters.