Work
What I have investigated, built and explained — grouped by what each piece demonstrates.
Research & evaluation
Original investigations with their evidence, results and stated limits — including the results that said no.
Same Dice, Different Geometry
Does a high overlap score mean the structure is right?
How matched Dice scores can conceal different structural errors in controlled vessel-mask surrogates.
My role First author, with Tudor Barbu (Romanian Academy).
Structural mechanics / numerical evaluation
Does an elegant geometry predict behaviour better than a simple baseline? It didn’t — and that result is reported.
Comparing behavioral distances in synthetic spring networks, including a negative result for the tested local geometry.
My role I designed, preregistered and ran the experiment, and wrote the report.
PDE-Refined Local Fractal Dimension Prior Conditioning and Topology-Aware Refinement for Retinal Vessel Segmentation with a Swin-UNet-Style Backbone
A published retinal vessel segmentation study, presented with its full citation and primary source.
RSKAN — interpretable Kolmogorov–Arnold networks
Can a network that refines its own function state recover the underlying law better than classical methods?
A preregistered research prototype with decision gates fixed before each run. The first two chapters ended in documented negative results — sparse regression did it more simply. In a later multi-tone signal family the learned model beat the best classical baseline on every seed, at about 43% of its compute — within three-tone cases only; it did not transfer across model order.
Virtual Cell Challenge 2026
How do human cells’ gene-expression profiles respond to genetic perturbations?
A competition entry predicting perturbation responses for 300 target genes from public Perturb-seq data, built as a hash-locked, independently verifiable pipeline. The final entry is submitted; scores are not yet public.
AI engineering & systems
Systems designed, built, tested and operated — from European institutional software to an AI-assisted platform.
AgroUnu
When a coding agent says “done”, is the task actually done?
A farm-management platform — field maps, crop records and purchase-to-dispatch traceability — built under an agent-evaluation process in which an independent reviewer can reject every change.
My role I design and build the platform and the agent-evaluation process around it.
Production systems for European institutions
Backend and database work on the European Commission’s grant-evaluation system (2019–2026) and the EEAS Security Portal (2018–2019); before that, numerical bridge-design software at Bentley Systems.
Interactive lab
Small, exact explainers. They demonstrate mathematical communication, not research results.
Same Dice, Different Geometry
Move the same number of pixels and watch what stays connected.
Right prediction. Wrong reason.
Break a shortcut, compare two exact Shapley explanations and test a distribution shift.
Vector Geometry
Cosine similarity and Euclidean distance answer different questions. Drag two vectors to see which.
Ongoing research
Current questions, kept separate from completed work.
Mechanistic interpretability
Is a model right for the right reasons?
An ongoing research direction in understanding how neural networks represent and process information.
My role My own research direction, alongside the PhD.
Research interests →Try it: right prediction, wrong reason →