My name is Niklas. I am currently doing a masters at ETH Zürich, building an AI consulting & development company called LupAi, doing research, and working on various fun projects.
I'm a Master's student in Computer Science at ETH Zürich, specialising in AI and machine intelligence. I graduated with honors from Saarland University (top 5%) and spent an awesome year at the University of Washington.
What drives me is the intersection of technical work and creative problem-solving. I've built apps used by real people, published research, won hackathons, and taught complex topics to hundreds of students.
The best solutions are both technically elegant and intuitively simple.
AI consulting & development for small and medium-sized businesses. Built RAG pipelines and LLM-driven financial document extraction systems, enabling SMBs to unlock insights from their unstructured data: bringing enterprise AI capabilities to teams that actually need them.
"We were drowning in old archives. LupAi built a solution tailored to our workflow, giving journalists fast access to accurate, source-backed information they can trust."
A social discovery platform that transforms how people share and find places. Instead of star ratings, YouKnow lets you search by vibe. For example, you can search "a cozy spot for a first date" or "the best hidden rooftop bar." Built collaborative recommendation maps, gamified exploration with XP levels, and community-curated location data available globally.
Won the Claude Builder Club 2026 Hackathon in Zurich. A competitive pricing intelligence platform that automatically matches a retailer's catalog to competitor products. It replaces days of manual research with a vector search + LLM reranking pipeline.
GitHub
Built in 24h at the Swiss AI Hackathon by ETH Entrepreneur Club. A post-loss administrative companion guiding relatives through German and Swiss bureaucracy after a death: insurance claims, inheritance, AI-generated correspondence, OCR document parsing.
GitHub
Custom U-Net implementation for automated segmentation of the mitral valve in cardiac ultrasound sequences. Trained on annotated echocardiography data, enabling accurate and efficient diagnostic workflows for cardiologists.
GitHub
Cryptographic protocols like MPC protect data privacy but blind-spot data poisoning. UTrace provides post-hoc auditing that attributes model integrity failures to responsible data owners — combining gradient similarity analysis with user-level unlearning, without compromising MPC privacy guarantees.
arXiv preprint
Developed regularized generalized additive models (RELGAM) for survival analysis, capturing nonlinear covariate effects in clinical data. Compared against classical methods like glmnet, achieving more expressive hazard function estimates for patient outcome prediction.
Built a learned prompting framework for sentiment classification using LLMs, combining a neural prompt selector (DistilRoBERTa) with a greedy prompt optimizer. Outperforms standard prompting strategies on accuracy, F1, and NMAE across sentiment benchmarks.
GitHub
Two threshold secret sharing schemes over infinite algebraic structures — continuous linear sharing over ℝ and approximate linear sharing over ℤ — motivated by RSA distributed exponentiation. Secrets are hidden under Gaussian distributed shares, with security and reconstruction analyzed rigorously. Saarland University, March 2024.