I research how to make language models smaller, faster and fairer by intervening on their structure: pruning, distillation and activation analysis.
Research
-
Fairness Pruning: Locating Demographic Bias in GLU-MLP Layers via Differential Activations
Identifies neuron circuits for demographic bias that are dissociable from general capabilities, retaining 99.49% of capability under targeted intervention.
- Fragile Knowledge, Robust Instruction-Following: The Width Pruning Dichotomy in Llama-3.2
- Exploring GLU Expansion Ratios: Structured Pruning in Llama-3.2 Models
Talks
Invited speaker, STAC Summit London (Apr 2025) · Speaker, Dev.BG Conference (May 2026)
Books
- Rearchitecting LLMs: Structural Techniques for Efficient Models
-
Large Language Models Projects
Used as a textbook at PES University, the University of Texas at Dallas (Jindal School of Management) and West Ukrainian National University.
Open source
-
OptiPfair
Structured pruning, knowledge distillation and bias visualization for LLMs.
-
Large Language Model Notebooks Course
Hands-on course on fine-tuning, DPO, pruning, RAG and evaluation. 1,800+ GitHub stars.
-
WizardSData
Synthetic conversation datasets and bias analysis.
Demos & models (Hugging Face)
- OptiPFair Bias Analyzer
- Fairness-pruning prompt pairs : fairness-pruning-pairs-en · fairness-pruning-pairs-es
- Companion models for Rearchitecting LLMs
- Articles
Writing
Contact
Open to research collaborations.