Research

Getting reliable intelligence out of constrained resources — first scarce labels, now scarce compute.

Current work

I work on vision–language and large language models — how they actually run on real hardware, and what architecture and quantisation choices cost in accuracy, throughput and energy.

The doctorate

Programme: JD-ICE — Joint Doctorate in Interactive and Cognitive Environments
Host lab: Elios Lab, DITEN, Università degli Studi di Genova
Partners: Universidad Carlos III de Madrid · Queen Mary University of London
Period: Nov 2025 – 2028 (expected)

Earlier: semi-supervised learning

Pursued in parallel with full-time industry practice, in collaboration with Islamic Azad University, South Tehran Branch. I originated Semi-CART, a distance-based sample-weighting scheme that lets decision-tree induction exploit unlabeled test-distribution structure through a weighted Gini/entropy split criterion, improving accuracy in the label-scarce regime. The same weighting principle then carried over to deep networks as SemiDeep. Both ship as installable Python packages, so the published results are reproducible rather than merely described.

Publications Code & benchmarks