Selected projects.

Optical sequencing platforms like Slide-Seq record where each transcript came from by imaging a slide of millions of barcoded beads over many cycles. Recovering the barcodes from those images is harder than it sounds: fluorescence bleeds between channels, phasing drifts, and beads light up unevenly.

Optocoder is a Python pipeline for the whole decoding step — aligning cycles, calling bases, correcting phasing and prephasing, and training a small ML model against Illumina ground truth where it exists. It was built around the Rajewsky Lab's wet-lab setup and published in NAR Bioinformatics.

Single-cell RNA-seq tells you what each cell is doing, but dissociation destroys the information about where it sat in the tissue. novoSpaRc reconstructs that spatial structure de novo, using optimal transport to find the arrangement of cells that best explains their expression similarity plus whatever partial spatial reference is available.

I co-led the 2021 Nature Protocols release, which focused on scalability — making the algorithm tractable on larger tissues and higher-resolution maps — and on a tutorial-driven API. The library has a steady user base (133★, 42 forks) and shows up in spatial-omics course material.

During a research visit at the Hebrew University of Jerusalem, I built a pipeline that uses a pre-trained protein language model (ProtBERT) to find candidate anti-insecticidal proteins directly from raw sequence data.

Most of the work went into the integration rather than the model: combining transformer embeddings with biological domain knowledge so the predictions could be read and prioritised by biologists instead of coming out as a black-box ranking.

Cell-based medicinal products have to be screened for particulate impurities. Flow imaging microscopy produces thousands of particle images per sample — too many to review by hand, and too varied for classical thresholding.

At Coriolis Pharma I built the deep-learning system for it: ingestion, training, evaluation, and a Python package around PyTorch/TensorFlow so scientists could retrain and audit the models themselves. The work was published in Cytotherapy with collaborators at Leiden University.

My first research work was real-time needle-tip localisation for an autonomous robotic biopsy system. The robot worked off a live ultrasound stream; my algorithms tracked the needle in that stream so the controller could keep it on target.

I wrote the C++/CUDA interface used during testing, did the Gabor-filter work for needle detection in 2D ultrasound, and contributed to five papers (ICRA 2016, ICRA 2017, ICAR 2015, Mechatronics 2019, and a 2020 TJEECS paper on curved-needle localisation).

CARTITUDE-1 is the phase 1b/2 study of ciltacabtagene autoleucel (cilta-cel) in relapsed/refractory multiple myeloma. I contributed to the biomarker work — integrating single-cell transcriptomic, proteomic, and clinical readouts to understand which patients respond and why.

The collaboration spans clinical, translational, and manufacturing teams; the manuscript is currently under revision at Blood Cancer Discovery.