SZStefano Zizzi

AI Engineering · Torino, Italy

Building AI systems forperception.language.trust.

I’m Stefano Zizzi, a Computer Engineering M.Sc. student specializing in Artificial Intelligence and Data Analytics, working where rigorous experimentation meets reliable software.

View selected work Download CV Get in touch
Selected research & engineering work

01 / Selected Work

Projects with evidence.

Each project is framed around the problem, technical approach, personal contribution and verifiable output.

01Research

Computer Vision · Anomaly Segmentation

MAAS — Road Scenes

A reproducible evaluation and analysis workspace for semantic-segmentation anomaly detection in road scenes.

Problem
Autonomous systems need to identify unfamiliar road objects while preserving reliable in-distribution segmentation.
Approach
Compared ERFNet and EoMT across common OOD benchmarks, calibration sweeps, boundary errors, object scale, depth bands, semantic confusions, resolution trade-offs and attention.
Contribution
Built the config-driven evaluation pipeline, fine-grained analysis tooling and an EoMT fine-tuning workflow with COCO outlier cutout augmentation.
Research output

A traceable benchmark with reusable evaluation wrappers, OOD metrics, calibration studies and generated analysis reports.

  • Python
  • PyTorch
  • EoMT
  • ERFNet
  • Cityscapes
View case study Explore the repository
02Complete

NLP · Political Discourse

CLARITY

Detection and classification of evasive political answers across transformer encoders, prompted LLMs and QLoRA fine-tuning.

Problem
Political answers can appear relevant while avoiding the question; the goal was to classify clarity beyond surface-level fluency.
Approach
Evaluated transformer encoders, prompted large language models and parameter-efficient QLoRA fine-tuning on the QEvasion task.
Contribution
Co-developed the experimental pipeline, model comparisons and reproducible project environment for the SemEval-style challenge.
Best mapped clarity

0.608 Macro-F1 on the best mapped-clarity configuration.

  • Transformers
  • LLMs
  • QLoRA
  • NLP
  • Docker
View case study Explore the repository
03Research

Explainable AI · Evaluation

EQE

An experimental framework for measuring whether local explanations remain faithful under a human-readable feature budget.

Problem
Standard XAI metrics can reward a complete explanation without testing whether its fidelity survives when users can inspect only a few features.
Approach
Introduced Complexity-Calibrated Local Concordance (CCC) and benchmarked SHAP, LIME and MAPLE over K = 5–9 features, two datasets, two models and multiple seeds.
Contribution
Implemented the benchmark orchestration, normalized additive contributions across explainers, developed the CCC metric and produced comparative visual analysis.
Main finding

SHAP delivered the strongest compact faithfulness; top-K features usually retained more signal than random-K controls.

  • XAI
  • SHAP
  • LIME
  • MAPLE
  • XGBoost
View case study Explore the repository
04In Development

Computer Vision · Multi-task Learning

Minerva

A modular Computer Vision backend for multi-attribute clothing classification.

Problem
Clothing understanding requires several related attributes to be predicted consistently from the same visual representation.
Approach
A shared ResNet-50 backbone feeds multiple task-specific prediction heads through a modular backend architecture.
Contribution
Designing the backend boundaries, shared feature extractor and extensible multi-head inference structure.
  • Python
  • PyTorch
  • ResNet-50
  • Computer Vision
View case study Follow development

02 / Technical Focus

Depth over buzzwords.

My work spans the model, evaluation and engineering layers—because useful AI depends on all three.

01perception

Computer Vision

Semantic segmentation, OOD detection and multi-task visual learning.

02language

NLP & LLMs

Transformer encoders, prompted models and parameter-efficient fine-tuning.

03trust

Explainable AI

Faithfulness, compact explanations and trustworthy evaluation.

04systems

ML Engineering

Reproducible experiments, clear interfaces and evidence-led delivery.

Working principle

“A strong result is not just a metric. It is a result you can explain, reproduce and interrogate.”

03 / Education & Experience

A foundation built in layers.

From Computer Science fundamentals to research-led AI engineering and applied machine learning.

  1. 01
    2024 — Present

    Politecnico di Torino

    M.Sc. in Computer Engineering — AI & Data Analytics

    Advanced study across Machine Learning, Deep Learning, Big Data systems and the mathematical foundations of AI.
  2. 02
    Academic experience

    University of Urbino Carlo Bo

    Human Activity Recognition internship

    Worked on an end-to-end Human Activity Recognition pipeline, connecting sensor data, experimental design and model evaluation.
  3. 03
    Previous

    University of Urbino Carlo Bo

    B.Sc. in Applied Computer Science

    Built the software-engineering and mathematical foundation that now supports research-oriented AI work.

04 / Open Source & Community

Knowledge compounds when it is shared.

Beyond research projects, I maintain and contribute to student-facing resources, peer collaboration and academic representation.

Explore InformaticaUniurb
20+GitHub stars
9Forks
Open notes, built together.A collaborative collection of Computer Science course notes created for University of Urbino students.

05 / Contact

Let’s build something
worth explaining.

I’m interested in research collaborations, internships and engineering work across Computer Vision, NLP and trustworthy AI.

Emailzizzi.stfn@gmail.comCodegithub.com/itsPinguizNetworklinkedin.com/in/zizzi-stefano