Things I make.

Apps, engineering builds, AI experiments and research, one collection, no artificial split.

AllAppsEngineeringAI / MLResearch

research2025–2026

Advanced Ceramic Thruster Development for Green Space Propulsion

Research project on the development of an additively manufactured ceramic thruster for water-based satellite propulsion. The work focused on a reliable digital and experimental development route from material and process preparation through DLP printing and thermal post-processing to characterization. In parallel, resincreator.de was built as a digital lab and process documentation system to structure material, manufacturing and test data and prepare them for future ML/AI-based recommendation systems.

mobile2026–now

Pulse

A private iOS app that turns Google Fitbit Air data, via Google's new Health API, into Whoop-style daily metrics: a recovery score from HRV, resting heart rate, sleep and respiration against a personal 30-day baseline, cardiovascular strain (0–21), sleep need and debt with a phase hypnogram, a baseline health monitor (resting HR, HRV, respiration, SpO₂, skin temperature), and a 'Pulse Age' biological-age estimate. Built natively in SwiftUI with a home-screen recovery-ring widget; all data stays on the device, no server, no subscription.

iOSSwiftUIGoogle Health APIHealth AnalyticsVisit

ai-ml2026–now

Legacy AI

Legacy AI is an AI-powered knowledge-preservation platform that captures the knowledge an organization actually runs on. A multi-tenant SaaS with role-based access, it combines AI voice interviews, automated knowledge extraction and semantic search to turn tacit expertise into a searchable, living knowledge base, so critical know-how isn't lost when key people leave.

Next.jsElevenLabspgvectorReactVisit

ai-ml2025–now

Reinforcement Learning & Behavioural Cloning

Analyzed a real-world robotics demonstration dataset for connector insertion in cable assembly. Performed time-series preprocessing, value estimation (discounted returns) and visual analytics of human search strategies. Trained a Behavioural Cloning model (Keras) to imitate expert actions from state signals (pose, wrench, twist) for safer, faster policy initialization.

PandasTensorFlowRoboticsHyperparameter TuningVisit

event2026

EHL Paris Research Hackathon

At the EHL Paris Research Hackathon, our team worked on the MSD Floor Plan Generation Challenge. The task was to generate plausible room layouts with room polygons and room types from empty apartment outlines. In 48 hours, we built an end-to-end pipeline for data preprocessing, layout generation, consistent rendering, evaluation with FID, Density and Coverage, and validation of the generated floor plans.

HackathonGenerative AIMachine LearningComputer VisionRead more Visit

research2026

Governing the Tools, Not the Agent: A Capability-Governed Deployment Platform for AI-Assisted Operations

Self-published paper on Nexus, my self-hostable deployment platform. Its thesis: expose every operation, to humans and AI agents alike, through one audited surface of capabilities, each tagged with a risk class, so an agent can observe and propose freely but never unilaterally perform an irreversible action without a human-issued, per-resource grant. The paper covers the capability model, the platform architecture and the zero-downtime deploy pipeline, validated on the three-machine fleet that runs this site.

AI AgentsDeploymentCapability GovernanceMCPVisit

app2026

CodeDraw — Code-to-Diagram Editor

CodeDraw is a browser-based diagram editor and AI-native render API. Describe diagrams in a minimal DSL (nodes, edges, labels) and get a live Excalidraw canvas instantly; export as PNG, SVG or JSON via HTTP. Integrates with ChatGPT, Claude, Cursor and VS Code Copilot through a Model Context Protocol server.

TypeScriptReactExcalidrawViteVisit

research2026

A Fact-Grounded, Version-Attributed Pipeline for AI-Assisted Job Applications

Self-published paper on Dossier, my self-hosted pipeline for AI-assisted job applications. Its thesis: career facts live once in a relational database, and every CV and cover letter is a disposable projection of it. Phrasings are stored apart from facts as versioned, bilingual, author-attributed texts, and an LLM collaborates through schema-validated MCP tools that let it rephrase, translate and propose, but never invent a fact. The paper covers the two-axis model, the architecture and the render pipeline, with a blind evaluation of the generated letters.

Job ApplicationsLLMMCPFact-GroundingVisit

event2026

TUM.ai x Unite Hackathon

At the TUM.ai x Unite Hackathon, our team built machine learning solutions for real-world procurement and recommendation problems. We worked on scalable feature normalization across large product catalogs and core demand forecasting to make product data more comparable and improve recommendation quality. The team achieved 1st place in Challenge 1 and 1st place in Challenge 2 Level 1.

HackathonMachine LearningProcurementFeature NormalizationRead more Visit

web2025–2026

AI Resin Creator

Companion platform to the ceramic-thruster research: a documentation and research tool for material- and process-data management, covering data collection, organisation and visualisation, built on a React/Supabase frontend with a structured database. An ML-based recommendation pipeline is in development, currently on hold pending larger datasets.

ReactSupabaseTypeScriptVisit

engineering2023–2024

3D-Printed Water-Jet Propelled Surfboard

Designed and built a modular, water-jet-propelled surfboard around a mostly 3D-printed propulsion unit (FDM/PLA). Integrated a waterproof electronics bay, a modular battery system, a high-power brushless motor and a water-cooled ESC, with jet geometry iterated in CAD. A hands-on additive-manufacturing and rapid-prototyping project.

Additive ManufacturingRapid PrototypingJet DriveCAD DesignRead more