CV

I'm mining alpha from alternative data for the prop business. Due to pretty strict confidentiality agreements, I cannot disclose much more than that.

Together with some amazing colleagues, I built out the multimodal ETL and synthetic data generation infra at Liquid AI. Most notably, this work led to the release of the LFM2-Audio-1.5B foundation model which achieved conversational quality that rivaled 10× larger models. This ultimately contributed to a multi-year partnership with Mercedes-Benz to bring embedded, on-device LFMs to MB vehicles across North America.

Having explored Large Language Models (LLMs) as they came out during my time in graduate school, I was keen to be a part of this emerging trend. I joined my former classmate as a founding engineer at Evidium to build evidence-grounded AI systems for healthcare. We focused on scalable architectures for reliably detecting, extracting, and reasoning on medical entities appearing in terabyte-scale data pulled from electronic health records. We patented, built and deployed a medical knowledge platform, currently in use at a sizeable regional U.S. hospital system.

I began my career at DB's sales and trading graduate program in Frankfurt, Germany. I rotated in correlation and xVA trading before joining the interest rate derivatives structuring team. As a structurer, my work involved development and pricing of exotic interest rate products for DACH area institutional clients. I later moved to DB's London office to focus on the EMEA corporates and governments structuring business.

During the pandemic and last part of my master's, I researched and implemented Temporal Difference (TD) Learning methods for a proprietary trader in collaboration with Dr. Daniel Bloch. This work was conducted under NDA, but our further research on sample-efficient learning of price densities and derivative payoffs using TD-λ methods is available to read on SSRN.

Alongside my studies, I actively participated in data science and finance competitions while organizing the TechLabs Barcelona program. Prior to my CEMS exchange at London School of Economics, I interned at Deutsche Bank in Frankfurt, Germany. My thesis was on the information content of S&P 500 index options. Our continued work on this was later published in The Journal of Derivatives.

After two years at the Vaasa campus, I moved to Ludwigshafen, Germany and wrote my thesis in collaboration with the credit risk department at BASF SE. Following that, I completed an exchange at Sophia University (上智大学) in Tokyo while working at Mitsubishi FUSO's material cost controlling department in Kanagawa, Japan.

We present LFM2, a family of Liquid Foundation Models (350M-8.3B parameters) designed for efficient on-device deployment. Using hardware-in-the-loop architecture search under edge latency and memory constraints, we obtain a compact hybrid backbone combining gated short convolutions with grouped query attention, delivering up to 2x faster prefill and decode on CPUs than similarly sized models. We also release multimodal and retrieval variants spanning vision-language, audio, and retrieval, all with open weights and deployment packages.

We propose option-implied measures of conditional dispersion and asymmetry based on quantiles and expectiles inferred from weekly options. All quantities are by construction forward looking and non-parametrically estimated with a novel arbitrage-free natural smoothing spline technique that produces quick and empirically accurate volatility smiles. We find that the proposed option implied indicators exhibit short, medium, and long-term predictive ability for the U.S. equity risk premium and market volatility, both in- and out-of-sample, and outperform equal indicators inferred from historical returns.

We extend Long Short Term Memory (LSTM) networks with parametric volatility model prediction heads that enable arbitrage-free forecasting and interpolation of the entire implied volatility surface dynamics within a single model. The approach is tested on out-of-sample daily SPX options data and shows strong forecasting power on liquid expiries.

We combine Temporal Difference learning (TD-λ) with the backpropagation (BP) algorithm to achieve sample efficient estimation of conditional expectations for path dependent events (across time & space). We benchmark our method and reach comparable results to Monte Carlo simulation with only two thousandth of the number of trajectories on a range of well known stochastic processes.

I present how we built large-scale synthetic data generation infrastructure on Ray at Liquid AI.

I discuss (in Swedish) use-cases of AI in the healthcare industry.

I present our paper Smiling in Action about dynamic implied volatility surface modeling using LSTMs.

I discuss my early career beginnings, and the exciting advancements happening in machine learning.