Research
I want to build AI that acts in the physical world and that people can check. My work runs on three threads. Each one ends in code that others can run.
Ich will KI bauen, die in der physischen Welt handelt und die Menschen prüfen können. Meine Arbeit folgt drei Linien. Jede endet in Code, den andere ausführen können.
Research threads
Embodied AI and robot learning
- Question
- How can a robot learn a control policy in simulation that transfers to real hardware?
- Approach
- RL in MuJoCo and ROS 2 with reward design, policy training and my own CUDA benchmarking code. Kiko at Barycenter Lab is our real robot test bed.
- Status
- In progress. MSc thesis with Dr. Fahmid Al Farid.
Multimodal health AI
- Question
- Can one model read brain, face, voice and language signals and explain what it found in plain words?
- Approach
- NeuroPrism reads EEG, facial behaviour, speech and interview text to find markers of depression, ADHD and related conditions across five public clinical datasets. It pretrains an EEG transformer on 26,000+ clinical recordings and fine-tunes it on 1,274 subjects. A grounded language model turns feature attributions into plain language reports. Earlier work: my 2024 IEEE paper on home stroke monitoring.
- Status
- In progress. Research project, not a medical device. Code release planned.
ML evaluation and benchmarking
- Question
- When is a gain real, and when is it luck?
- Approach
- A 23-model forecasting benchmark with statistical tests and KKT boundary theory.
- Status
- Manuscript in preparation. Code is public.
Publications
Manuscripts in preparation
-
Convex Forecast Combination and Sparsity via KKT Boundaries: Multi-Dataset Benchmarking in Energy Time Series
Manuscript in preparation. My role: First author.
We benchmark 23 forecasting models on 5 energy datasets under 2 protocols, with rolling origin horizons 1, 3, 6, 12 and 24 over 24 origins. We test the results with Diebold-Mariano, Cohen's d, Cliff's delta, and Friedman with Nemenyi. KKT boundary theory explains when the convex ensemble collapses onto tree learners.
Wir vergleichen 23 Prognosemodelle auf 5 Energiedatensätzen unter 2 Protokollen, mit rollierenden Horizonten 1, 3, 6, 12 und 24 über 24 Ursprünge. Wir prüfen die Ergebnisse mit Diebold-Mariano, Cohens d, Cliffs Delta sowie Friedman mit Nemenyi. Die Theorie der KKT-Grenzen erklärt, wann das konvexe Ensemble auf die Baummodelle zusammenfällt.
Peer reviewed
-
Continuous Health Monitoring for Early Stroke Prediction at Home
IEEE ICEECT, 2024. My role: First author.
A home system that flags early stroke risk from continuous body signals with machine learning.
Ein System für zu Hause, das ein frühes Schlaganfallrisiko aus laufenden Körpersignalen mit Machine Learning erkennt.
-
An Effective Approaches for Auto Bill Generator System
IEEE ICEECT, 2024. My role: Co-author.
A billing system for shops. A camera model finds each item and weight sensors check it.
Ein Kassensystem für Geschäfte. Ein Kameramodell erkennt jeden Artikel, und Gewichtssensoren prüfen ihn.
Profiles
Google Scholar and ORCID.