Teaching machines to read images — and write about them.
I research AI for medical imaging at the Pattern Recognition Lab, Friedrich-Alexander-Universität Erlangen-Nürnberg — where computer vision meets large language models. Before the PhD I spent five years building production ML: face recognition and anti-spoofing systems at Gaze and CPSD Technologies in Dhaka, then data and LLM work at the BMW Group and Innomotics (Siemens) in Bavaria.
Multimodal AI
Nürnberg, DE
My thread through all of it: getting models to reason across more than one modality at a time, and making them dependable enough to leave the notebook.
Vision–language models that read a scan and produce clinically faithful text. My MSc thesis generated region-guided clinical reports from chest X-rays using LLMs — the PhD at the Pattern Recognition Lab carries that line of work forward.
Retrieval-augmented systems that stay grounded in their sources. At Innomotics I built RAG applications over dense industrial documentation — evaluation, chunking strategy and hallucination control included.
Face anti-spoofing, detection and landmark models small enough to run on a phone. Three of my papers come from this area, and I shipped it for real — a facial attendance system in production, 3× faster on edge devices.
A few things I've built end to end, from research code to deployable artefacts.
A sub-1 MB face detector with 5-point landmarks, built for mobile and edge deployment across ncnn, ONNX and PyTorch.
Read more
An end-to-end ELT pipeline and analysis of Germany's EV charging infrastructure — regional gaps, capacity and where to build next.
Read more
A sequence-to-sequence ASR system for Bangla in PyTorch, with the full data, training and validation toolchain around it.
Read moreICPR 2022 — 26th International Conference on Pattern Recognition
Sensors (MDPI) · 21(8), 2799
I'm always up for conversations about vision–language models, medical AI, collaborations, or a good supervision question. My inbox is open.