Research & Papers

Peer-reviewed work on efficient AI — Bayesian sparsification, quantization, and deployment-robust model compression.

2026 · Under Review

Efficient Image Enhancement on the Edge

Submitted to NeurIPS 2026 (first author)

First-author work on state-of-the-art image enhancement designed for edge deployment, conducted within the INSAIT / ETH Zurich research ecosystem in collaboration with academic and industry partners. Details to follow upon publication.

2025

Quant-Trim in Practice: Improved Cross-Platform Low-Bit Deployment on Edge NPUs

EurIPS 2025

A training-phase method producing hardware-neutral checkpoints robust to backend and precision choices. Combines progressive fake quantization with reverse pruning to tame outlier-driven scale inflation. Agnostic to quantization schemes (symmetric/asymmetric, per-tensor/per-channel, INT8/INT4), requires no vendor-specific graph changes, and avoids per-backend retraining.

2024

Shaving Weights with Occam's Razor: Bayesian Sparsification for Neural Networks Using the Marginal Likelihood

NeurIPS 2024 (Main Track) & AABI 2024

A principled Bayesian approach to neural network sparsification using the marginal likelihood, achieving state-of-the-art compression while maintaining model performance — with applications to vision transformers and LLMs.

2023

From Judgement's Premises Towards Key Points

arXiv

Advanced NLP methods for extracting key points from legal texts, enhancing automated document analysis with BERT, IBM Debater, and graph-based approaches.

Reviewing

Scientific reviewer for top-tier machine learning conferences (ICML, NeurIPS) and journals (JAIR), covering probabilistic inference, compression, efficient AI, computer vision, and LLMs.