MSc student in Artificial Intelligence Systems at the University of Trento, with a BSc in Computer Science from the University of Camerino (104/110) and industry experience in digital transformation consulting.

I've worked on LLM agent architectures & evaluation, vision-language models, and probabilistic retrieval systems. I care about building AI systems with statistical rigor, measurable benchmark performance and cost-efficient trade-offs.

Currently seeking AI/ML or Software Engineering roles & internships (fully flexible on setup, ready to relocate).

LinkedIn · Email · GitHub

davideCabitz/Deep-Learning-Project · MSc AI Systems, UniTrento

Engineered a PyTorch multi-modal image retrieval pipeline conditioning frozen CLIP (ViT-B/32) embeddings with complex multi-attribute constraints to solve geometric failure modes in state-of-the-art CVPR baselines.

• Algorithmic Innovation: Developed A-BAP (Amortized Bayesian Attribute-Posterior scoring), combining 40 calibrated logistic attribute probes, a Poisson-binomial dynamic program for exact identity preservation, Chow-Liu dependency trees, and ListNet listwise ranking loss.
• Results: Achieved $R@10 = 0.413$ ($4\times$ gain over vanilla CLIP and $+61%$ over non-probabilistic models, $p < 0.001$, Holm–Bonferroni corrected) on CelebA. Broke the $0%$ retrieval floor on pure-negation queries (-Male, -Mustache), raising $R@5$ from $0.000$ to $0.222$.
• Stack: Python, PyTorch, CLIP (ViT-B/32), Hugging Face Transformers

davideCabitz/AngryAgents · Designing Large Scale AI Systems, UniTrento

Engineered a multi-agent web application for real-time group chat simulations across 100+ LLM-scraped AI persona agents, backed by an automated multi-LLM evaluation pipeline.

• Statistical Rigor & Judge Pipeline: Validated judge agreement using Cohen's Kappa ($\kappa = 0.66$, "Good" reliability) and achieved a $4.83/5.0$ mean persona fidelity score across 21 multi-turn benchmark chats ($95%$ CI $[4.74, 4.93]$).
• Model & Cost Tradeoff: Quantified trade-offs between models, demonstrating that upgrading to flagship models raised character accuracy from $53.8%$ (Macro F1: $0.656$) to $97.5%$ (Macro F1: $0.975$, $p < 0.001$) at a $17\times$ cost multiplier.
• Stack: Python, OpenAI API, LLMs, RAG, SciPy

davideCabitz/AMS---Autonous-Software-Agents-for-Deliveroo.js · Autonomous Software Agents, UniTrento

Engineered two cooperating autonomous BDI agents competing in a grid game that dynamically select among 8 strategies to optimize time-decaying parcel delivery.

• Architecture: Combined $A^*$ graph navigation, PDDL planning for spatial puzzles, and probabilistic opponent modeling with a decay-aware value function.
• LLM Integration: Built a ReAct GPT-4o command layer with ~40 tools to parse and cost-price natural language missions, alongside a custom JSON coordination protocol enabling inter-agent parcel handoffs.
• Stack: JavaScript (Node.js), OpenAI SDK, LiteLLM, PDDL, $A^*$ Search

Junior Consultant Intern · Business Integration Partners (BIP) S.p.A. (Milano, Italy) · May 2025 – Sep 2025 Supported public sector digital transformation and modernization initiatives under the National Recovery and Resilience Plan (NRRP / PNRR). Mapped AS-IS/TO-BE processes, conducted data analysis and KPI tracking for executive performance monitoring, and managed cross-functional stakeholder coordination.

• AI/ML & Vision: PyTorch, CLIP (ViT-B/32), Hugging Face, RAG, Probabilistic Modeling
• LLMs & Multi-Agent Systems: OpenAI API, BDI Agents, ReAct Tooling (~40 tools), PDDL Planning, LiteLLM, Automated Multi-LLM Judge Pipelines
• Languages: Python, JavaScript (Node.js), C++, Java, SQL
• Languages Spoken: English (IELTS Academic 7.0 / C1), Japanese (JLPT N4 / B1 equivalent), Italian (Native)