Building intelligent systems that connect AI research with real-world impact.
π Originally from Ethiopia
I hold a BSc in Computer Science and Engineering from Mekelle University, Ethiopia, graduating with a 3.96/4.00 CGPA and ranking in the top 1% of 800+ engineering students. I bring over four years of experience spanning software engineering, data analytics, and machine learning.
My interests lie at the intersection of Computer Vision, Medical AI, Deep Learning, Explainable AI, and AI Engineering. I enjoy designing experiments, comparing model architectures, analyzing model behavior, and translating technical ideas into practical applications.
I believe effective AI development requires more than achieving a high accuracy score. It requires thoughtful evaluation, reliable data pipelines, understanding failure cases, and building systems that are useful beyond a research notebook.
- π¬ Computer Vision: Image classification, CNNs, Vision Transformers, transfer learning, and visual explainability.
- π₯ Medical AI: Retinal OCT classification, medical image analysis, and healthcare decision-support prototypes.
- π§ Machine Learning: Supervised and unsupervised learning, ensemble methods, statistical analysis, and model evaluation.
- βοΈ Efficient AI: Parameter-efficient fine-tuning, LoRA, hyperparameter optimization, and model comparison.
- π Data Science: Exploratory data analysis, feature engineering, clustering, visualization, and interpretable models.
- π» AI Engineering: Python applications, REST APIs, databases, backend development, and deployment.
- π AI for Impact: Healthcare, environmental applications, cultural heritage digitization, and technology access.
Focus: Deep learning for retinal imaging and multi-class medical image classification.
VisionCare AI explores deep learning approaches for classifying retinal optical coherence tomography (OCT) scans into four categories:
- Choroidal neovascularization (CNV)
- Diabetic macular edema (DME)
- Drusen
- Normal
Key contributions and reported results
- Worked with a dataset of 83,605 OCT scans.
- Fine-tuned four CNN architectures: VGG16, InceptionV3, ConvNeXt-Tiny, and EfficientNetV2-S.
- Explored soft-voting ensembles and XGBoost stacking.
- Achieved 97.08% accuracy and F1 with the reported ensemble approach.
- Addressed class imbalance using class-weighted loss, weighted sampling, and label smoothing.
- Evaluated per-class F1 scores to examine differences in model performance across categories.
- Used differential learning rates, OneCycleLR, and mixed-precision training.
Technologies
Python PyTorch CNNs Transfer Learning XGBoost Ensemble Learning
Why it matters: Medical imaging models should be evaluated across disease categories, not only by overall accuracy. This project explores model comparison, class imbalance, and complementary ensemble predictions.
Focus: Computer vision, Vision Transformers, parameter-efficient fine-tuning, and visual explainability.
This project investigates image classification across 21 land-use categories using 2,100 aerial images from the UC Merced dataset.
Key contributions and reported results
- Compared a custom residual-attention CNN, DenseNet-121, and ViT-B/16 adapted using LoRA.
- Achieved 99.05% test accuracy with ViT-B/16 + LoRA in the reported experiment.
- Evaluated DenseNet-121, which achieved 96.67% accuracy in the reported experiment.
- Used perceptual-hash deduplication and stratified splitting to help reduce data leakage risks.
- Applied Optuna for hyperparameter optimization.
- Used macro-F1 alongside accuracy to evaluate classification performance.
- Applied Grad-CAM to inspect model predictions.
- Built a Streamlit dashboard supporting single-image and batch inference.
Technologies
Python PyTorch Vision Transformers LoRA Optuna Grad-CAM Streamlit
Why it matters: This project explores how parameter-efficient adaptation and careful evaluation can be used for visual recognition tasks, while making model predictions accessible through an interactive application.
Focus: Unsupervised learning, Bayesian networks, and interpretable machine learning.
This project explores patterns in a dataset containing approximately 413,000 records and 16 variables, combining clustering, probabilistic modeling, and explainability.
Key contributions and reported results
- Compared six clustering methods: K-Means++, K-Prototypes, K-Modes, agglomerative clustering, DBSCAN, and Gaussian mixture models.
- Achieved reported cluster purity between 69.6% and 73.6%.
- Used Gower distance to work with mixed categorical and numerical features.
- Applied PCA and t-SNE to visualize data structure.
- Learned Bayesian network structures using BIC, AIC, and domain constraints.
- Reported 68.8% held-out accuracy for the Bayesian network modeling task.
- Applied SHAP and a decision-tree surrogate to help interpret model behavior.
Technologies
Python scikit-learn Bayesian Networks SHAP PCA t-SNE Statistical Analysis
Why it matters: The project emphasizes understanding data structure and interpreting model behavior instead of relying exclusively on predictive metrics.
Focus: Machine learning for healthcare decision support.
Afya is a prototype that explores predicting risk for diabetes, malaria, and heart disease from patient-history data.
Key contributions
- Processed patient data and performed feature selection.
- Trained and evaluated multiple machine learning models.
- Tested the prototype on 1,000 patient records.
- Integrated the prediction workflow into a web-based application using Flask and Django.
- Explored how data-driven decision-support tools could be made accessible in settings with limited specialist care.
Technologies
Python scikit-learn Flask Django Machine Learning
Why it matters: The project connects machine learning workflows with software development to explore practical healthcare applications.
This is a prototype, not a clinically validated diagnostic system.
Focus: Optical character recognition and cultural heritage digitization.
This project explores OCR for handwritten manuscripts in Ge'ez, Amharic, and Tigrigna, including historical religious, scientific, and medical texts.
Key contributions
- Developed and evaluated an OCR pipeline using more than 150 manuscript pages.
- Reported 87% accuracy in the project.
- Used image-processing techniques to support manuscript recognition.
- Explored digital access to fragile historical documents.
Technologies
Python OpenCV TensorFlow OCR Image Processing
Why it matters: Language technology and document digitization can help preserve historical knowledge and make valuable manuscript collections easier to search and study.
My experience extends beyond model development into production-oriented software, analytics, and backend systems.
- Analyzed sales, customer, website, and marketing KPIs.
- Produced reports and recommendations for management.
- Contributed to reported improvements of 40% in web traffic and 25% in conversion.
- Automated recurring reports, reducing manual processing by 30%.
- Maintained databases with attention to data quality and integrity.
- Applied statistical and machine learning methods to large datasets.
- Produced more than 10 predictive models and analytical insights.
- Built interactive Python visualizations that improved data interpretation and reporting efficiency.
- Developed web applications using Django and React.js.
- Worked on API integrations and application performance.
- Contributed to a reported 35% reduction in API integration time and 26% improvement in page-load time.
- Built REST endpoints using Java and Spring Boot.
- Delivered more than 20 REST endpoints.
- Optimized database queries from approximately three seconds to under 500 milliseconds.
- Improved average API response time by 30%.
These experiences shape my approach to AI: I value clean implementations, efficient data handling, maintainable APIs, and practical delivery.
Methods: classification, clustering, CNNs, transfer learning, Vision Transformers, LoRA, ensemble learning, hyperparameter optimization, statistical analysis, and model evaluation.
Additional experience: REST APIs, Swagger/OpenAPI, Node.js, MySQL, PostgreSQL, database management, and full-stack application development.
Relevant areas of I am exploring noe includes:
- Machine learning and deep learning
- Computer vision and AI for signal/image processing
- Signal and imaging acquisition and modeling in healthcare
- Environmental sensing and imaging
- Advanced computational techniques for large imaging and signal datasets
- Statistics, mathematics, and data privacy
Mekelle Institute of Technology, Mekelle University
Mekelle, Ethiopia | September 2016 β October 2021
- CGPA: 3.96/4.00
- Graduated with Very Great Distinction
- Ranked in the top 1% of 800+ engineering students
- Achieved a perfect 4.00 GPA in six of ten semesters
A2SV Community Education β Selected for training in Python, data structures, and algorithm design and analysis.
I believe technical growth is even more meaningful when knowledge and opportunities are shared.
- Global Mentorship Initiative: Volunteer career mentor supporting mentees with career planning, interview preparation, and professional development.
- YES Ethiopia: Volunteer career mentor supporting students in exploring educational and professional opportunities.
- Student mentoring and tutoring: Supported students through tutoring and mentorship activities in Tigray.
- Black in AI: Member and selected participant in the Emerging Leaders in AI graduate program in 2026.
- AI community: Interested in contributing to a stronger, more inclusive AI ecosystem, particularly across Africa and underserved communities.
I am especially interested in exploring:
- Medical image analysis: Deep learning for retinal imaging and other healthcare imaging modalities.
- Label-efficient learning: Learning useful representations when labeled medical data is limited.
- Parameter-efficient fine-tuning: LoRA and related methods for adapting pretrained models.
- Generalization and robustness: Understanding how models behave across datasets and imaging conditions.
- Explainable AI: Using tools such as Grad-CAM and SHAP to inspect model behavior and failure cases.
- Computer vision for the environment: Image-based analysis for environmental monitoring and land-use applications.
- AI engineering: Building maintainable applications that integrate models, APIs, databases, and retrieval systems.
I'm interested in research that combines sound experimental design, reproducible evaluation, and meaningful real-world applications.
A learning repository covering machine learning and deep learning topics, including classical algorithms, ensemble methods, CNNs, RNNs, object detection, Transformers, and self-supervised learning.
A structured collection of Python learning materials and exercises covering data structures, algorithms, complexity analysis, and competitive programming concepts.
Explore more projects in my GitHub repositories.
I'm happy to connect with AI researchers, software engineers, students, and organizations working on meaningful technology.
- πΌ LinkedIn: linkedin.com/in/nigushaile
- π§ Email: n.haile@campus.unimib.it
- π GitHub: github.com/NigusHaile
If you are interested in computer vision, medical AI, machine learning, or building practical AI applications, feel free to reach out.
