Antonio Colamartino
Software Engineer
Machine Learning · NLP · LLMs · RAG · Agents · LangChain · PyTorch · TensorFlow · Azure · Terraform
Experience
Work
March 2025 - June 2026
Software Engineer - Analyst
Deloitte Nexthub, Bari
Software development with .NET and C# for scalable architectures. Maintenance and development of Azure resources.
Education
2026 - Ongoing
Master's Degree in Computer Science (LM-18)
University of Bari
Master's programme with a focus on Artificial Intelligence and Machine Learning.
2022 - 2026
Bachelor's Degree in Computer Science (L-31)
University of Bari
Graduated with a score of 106/110. Degree program focused on software development, algorithms, data structures, and machine learning.
2017 - 2022
High School Diploma in Computer Science and Telecommunications
I.I.S.S. G. Ferraris
Specialization in programming, computer networks, databases, and information systems.
Projects
Multi-Method XAI for Diabetic Retinopathy
Multi-Method Rule Extraction from Deep Learning for Interpretable Diabetic Retinopathy Grading
A clinical decision support system that detects diabetic retinopathy in fundus photographs and explains the reasoning behind each diagnosis with readable rules. Developed as a thesis project, it combines the performance of deep neural networks with the interpretability of extracted rules.
Architecture EfficientNet-B5 teacher CNN with Focal Loss and label smoothing; explainability via Grad-CAM++; rule extraction with Decision Tree, LIME and RIPPER; hybrid CNN + rules system.
- amicroaneurysms - Grad-CAM++ saliency 0.91
- bintraretinal haemorrhages - Grad-CAM++ saliency 0.84
- chard exudates - Grad-CAM++ saliency 0.78
- doptic disc
- emacula and fovea
ICDR severity grades (0-4)
Figure notes
- EfficientNet-B5 teacher CNN (~30M parameters) trained with Focal Loss and label smoothing (epsilon 0.1).
- Grad-CAM++ saliency maps quantitatively validated: Deletion/Insertion AUC and IoU against the annotated lesions of the DDR dataset.
- Interpretable rule extraction with three methods compared: Decision Tree distillation, LIME, RIPPER one-vs-rest.
- Hybrid CNN + rules system: post-hoc explanation, rule-guided prediction and weighted ensemble.
- 82.95% accuracy (quadratic weighted kappa 0.829) on the internal test set; post-hoc explanation covers 88.2% of predictions.
| Parameter | Value |
|---|---|
| Dataset | APTOS 2019 + EyePACS 2015: 37,933 fundus images, 80/10/10 split; XAI validation on DDR (12,522 images with pixel-level annotations) |
| Classes | 5 severity grades (ICDR scale 0-4) |
| Backbone | EfficientNet-B5: teacher CNN, ~30M parameters |
| Training | Focal Loss + label smoothing (epsilon 0.1) |
| Explainability | Grad-CAM++, Decision Tree, LIME, RIPPER |
| Stack | Python, PyTorch, scikit-learn, Jupyter |
Source code and documentation on GitHub
Paper in preparation: it will be available here once published.