portfolio.sysnode 20 · turbopack · 64-bit
0%
JacopoParretti
Verona--:--
Available for Work
MSc · Artificial Intelligence
Jacopo
PARRETTI
AI Engineer  ·  Developer  ·  Researcher

Machine learning and software engineering, currently focused on language models. Pursuing an MSc in Artificial Intelligence at the University of Verona.

View Work
Verona, Italy
01 — About

About

·“I came to computer science through biology. Three years with genomic data taught me to reason about noise, dimensionality, and the gap between a clean experiment and a real one — habits that carry into everything I build now.”

My work covers information retrieval, supervised classification, LLM fine-tuning, and reinforcement learning. I care about understanding why a system works, not only that it does, and about writing code that holds up outside the notebook.

Alongside the MSc I work as a software engineer at North Plastik S.p.A., as project manager and specialist for AI/ML projects. I also contribute to open-source on GitHub — mostly ML and developer tooling.

Jacopo Parretti

Current interests

Large Language Models·Bioinformatics·Reinforcement Learning·NLP·Statistical ML·Open Source

Currently
  • MSc Artificial IntelligenceUniversità di Verona
  • Software EngineerNorth Plastik S.p.A.
  • Open-Source ContributorGitHub

In numbers

0Projects
0+Years coding
0Degrees
—Repos

Contact

02 — Projects

Selected

Projects.

I.

Uncertainty-Aware Model-Based RL for Autonomous Driving

A PyTorch reimplementation of MACURA (Frauenknecht et al., ICML 2024) — a model-based reinforcement-learning method that scales the length of imagined rollouts to the world model's own uncertainty — applied to the MetaDrive driving simulator. Benchmarks model-free SAC against model-based MBPO and MACURA, and ablates the uncertainty threshold that decides how far the agent trusts its model before truncating a rollout.

Python·PyTorch·Reinforcement Learning·Model-Based RL·MetaDrive
U
II.

CNN vs Vision Transformer — Where Does the Model Look?

A comparative study of CNN and Vision Transformer classifiers on the fine-grained Oxford-IIIT Pets dataset — asking not only which model is more accurate but where each one looks. Five variants are trained and compared, from ResNet-50 to a fine-tuned ViT-B/16, a ViT-S/16 from scratch, DeiT-S/16, and a frozen linear probe, with saliency surfaced through Grad-CAM and attention rollout. ViT-B/16 tops the test set at 93.7%; a transformer trained from scratch collapses to 22.7%, isolating its data-hunger.

Python·PyTorch·Vision Transformers·CNN·Grad-CAM·TensorBoard
C
III.

An E-Commerce Ontology in Description Logic

An OWL 2.0 ontology formalising the e-commerce domain — products, customers, orders, payments, reviews, categories and shopping carts — within the ALCQI description-logic fragment. Authored in Protégé and machine-verified for consistency and coherence with the HermiT reasoner, with design notes documenting every modelling decision.

OWL 2.0·Description Logic·Protégé·HermiT·Knowledge Representation
A
IV.

Comparative Architectures for LLM-Based Email Triage

Evaluates three routing strategies — frozen zero-shot prompting, LoRA fine-tuning, and DistilBERT classification — for automatic assignment of customer-support emails across five departments. LoRA reaches 70.4% accuracy with under 1% trainable parameters; DistilBERT peaks at 76.6%.

Python·PyTorch·LoRA·DistilBERT·HuggingFace
C
V.

Semantic Document Retrieval on the Reuters Corpus

TF-IDF vectorisation and cosine similarity pipeline for large-scale document matching across ~10,800 Reuters news articles. Accepts raw text or file input and returns ranked results with configurable percentile thresholds — no stopword removal, operating on raw token distributions.

Python·NLTK·scikit-learn·TF-IDF·Cosine Similarity
S
VI.

Multi-Class Nutritional Grade Prediction via Classical ML

Five-way Nutri-Score classification (A–E) trained on 250,000 Open Food Facts products. Benchmarks Logistic Regression, KNN, SVM, Random Forest, and XGBoost across a full ML pipeline — imputation, outlier removal, PCA, stratified splits. SVM tops validation at 83.5% accuracy.

Python·scikit-learn·XGBoost·PCA·Open Food Facts
M
VII.

A Decision Procedure for Equality, Lists, and Arrays

Satisfiability solver implementing congruence closure for the quantifier-free first-order theories T_E, T_cons, and T_A. Follows Bradley & Manna (2007), uses the largest-ccpar heuristic, and supports SMT-LIB 2.0. 571 JUnit tests, all passing.

Java·SMT Solving·Congruence Closure·Formal Verification·JUnit
A
VIII.

Reinforcement Learning for Smart-Grid Energy Optimisation

A research framework comparing three reinforcement-learning algorithms — SAC, PPO and TD3 — for energy optimisation in smart buildings, built on the CityLearn simulation environment. A modular, extensible architecture scales across multiple buildings and runs reproducible experiments behind a fair, shared evaluation harness.

Python·Stable-Baselines3·CityLearn·SAC·PPO·TD3
R
IX.

An RL Agent for the AIcrowd CityLearn Challenge

A reinforcement-learning setup targeting the KPIs of the AIcrowd CityLearn challenge — a custom Gym environment, reward-shaping utilities, and a two-phase PPO training pipeline on Stable-Baselines3. Ships with TensorBoard logging and tooling to extract and analyse scalar metrics straight from the event files.

Python·Stable-Baselines3·Gymnasium·PPO·CityLearn·TensorBoard
A
03 — Stack

The

Stack.

ML & AI00
Languages00
Web & APIs00
Infra & DevOps00
I.

ML & AI

5 tools
PyTorch
TensorFlow
●JAX
Hugging Face
scikit-learn
II.

Languages

4 tools
Python
TypeScript
Go
Rust
III.

Web & APIs

5 tools
Next.js
React
Tailwind CSS
FastAPI
Node.js
IV.

Infra & DevOps

6 tools
Docker
Kubernetes
PostgreSQL
Redis
Git
●AWS
04 — Education

Education

Diploma IIn Progress
2025 – Present
In progress

MSc Artificial Intelligence

Università degli Studi di Verona · Verona, Italy

Advanced studies in machine learning, probabilistic AI, and deep learning. Research oriented toward large language models, neural architecture design, and efficient inference.

Machine LearningDeep LearningNLPProbabilistic AILLMsNeural Architectures
Diploma IICompleted
2021 – 2024

Final mark

0

/ 110

BSc Bioinformatics

Università degli Studi di Verona · Verona, Italy

Three-year degree combining computational biology, statistics, and software engineering. Strong foundation in algorithms, genomic data analysis, and scientific computing.

AlgorithmsGenomicsStatistical AnalysisScientific ComputingPythonR
05 — Timeline

Across

five years.

2026 – Present

Work

Software Engineer

North Plastik S.p.A.

Project manager and specialist for AI/ML projects.

2025 – Present

Study

MSc Artificial Intelligence

Università degli Studi di Verona

Advanced studies in machine learning, probabilistic AI, and deep learning. Research focus on large language models and neural architectures.

2023 – 2026

Work

Sales Assistant & Data Analyst

JD Sports & Fashion

Retail floor work alongside internal analytics — trend reporting, KPI dashboards, and inventory insight.

2022 – Present

Side

Open-Source Contributor

GitHub

Contributions across ML and developer tooling repositories — bug fixes, feature PRs, and documentation.

2022 – 2023

Work

Sales Assistant

Foot Locker

First job. High-traffic retail floor — customer service, shift coordination, and learning how to work under pressure.

2021 – 2024

Study

BSc Bioinformatics

Università degli Studi di Verona

Three-year degree spanning biology, computer science, and statistics. Strong foundation in algorithms, data analysis, and scientific computing.

06 — Contact

Get

in touch.

Compose a message
or reach directly
GitHubLinkedInInstagram
Jacopo ParrettiAI Engineer · Verona, Italy© 2026 — Built with Next.js