Antonio Sabbatella AntonioSabbatellaUni

Antonio Sabbatella

MSc Data Science (110L/110) | AI Research Engineer

Specializing in Bayesian Optimization, Multi-Agent Systems, and Efficient LLM Architectures.
Bridging the gap between theoretical research and production-grade engineering.

🔬 Research & Engineering Focus

I focus on reducing computational overhead and automating complex reasoning in Generative AI systems. My work spans from architectural optimization to high-level agentic orchestration.

Deep Learning & Systems: Designing scalable architectures for DeepSeek Sparse Attention, RAG pipelines, and custom execution environments for agentic workflows.
Bayesian Optimization: Automating prompt engineering and multi-agent team composition (MALBO, BOInG) using Multi-Objective strategies.
Efficient NLP: Exploring Context Compression frameworks ($84%$ token reduction) and fine-tuning strategies to optimize inference costs on constrained budgets.

🛠 Tech Stack

Core:

LLM & Agents:

Infrastructure:

🏆 Featured Work

Project	Domain	Impact
LUDUS	Deep Learning / Kernels	Implementation of DeepSeek Sparse Attention for Qwen models. Reduces complexity to $O(N \cdot K)$. Optimized for consumer hardware (T4).
MALBO	Multi-Agent / Bayesian Opt	Pareto-Efficient Multi-Agent Optimization. Finds optimal trade-offs between Cost and Performance for agent teams using Multi-Objective Bayesian Optimization. Features a custom fork of `smolagents` for heterogeneous LLM swapping.
StudyWithWisp	Full Stack AI / SaaS	AI Platform for student prep (Flashcards/Simulations). Built with Next.js & Python on GCP. Scalable RAG pipeline with Prisma. (Private Repo)
UiNav	Computer Vision / Agents	Autonomous UI interaction system combining YOLO (finetuned) with LLMs for natural language-driven browser automation.

📄 Selected Publications

[High Impact] Prompt optimization in large language models A. Sabbatella, A. Ponti, I. Giordani, A. Candelieri, F. Archetti (2024)
Mathematics 12 (6), 929
Seminal work on Bayesian strategies for prompt engineering. Cited by 50+ researchers.
MALBO: Optimizing LLM-Based Multi-Agent Teams via Multi-Objective Bayesian Optimization Antonio Sabbatella (2025)
arXiv preprint arXiv:2511.11788
Framework for identifying Pareto-efficient agent teams, achieving 65.8% cost reduction vs baselines.
Bayesian Optimization for Instruction Generation A. Sabbatella, et al. (2024)
Applied Sciences 14 (24), 11865
Introduction of the BOInG framework, reducing GPU memory requirements by two orders of magnitude.
Bayesian Optimization Using Simulation-Based Multiple Information Sources A. Sabbatella, et al. (2024)
Machine Learning and Knowledge Extraction 6 (4)
Advanced combinatorial optimization using multi-source information fusion.