I'm an AI Researcher at Baku State University and an MSc student in Knowledge Acquisition Systems at Azerbaijan Technical University. I earned my BSc in Computer Science at Baku State University and am completing a dual degree with Holon Institute of Technology (Israel). I work where mathematics meets applied AI: I like problems where a clean mathematical formulation tells you what a model can and cannot learn.
- 🔬 AI Researcher @ Baku State University
- 🏆 ICPC Coach @ Baku State University
- 👨🏫 C++ & Algorithms Instructor @ Codemy
- 🎓 MSc in Knowledge Acquisition Systems @ Azerbaijan Technical University
- 🎓 BSc Computer Science @ Baku State University, dual degree with Holon Institute of Technology (GPA 93.21)
- 🇦🇿 Building AZ-Eval, an open parallel Azerbaijani-English LLM benchmark
- 🎯 Open to research collaborations, AI/ML engineering roles and conference connections
- 📫 Reach me at nihatgaribli25@gmail.com
- 🧾 ORCID: 0009-0003-4755-6356
| Project | What it does | Stack |
|---|---|---|
| AZ-Eval | Parallel Azerbaijani-English LLM benchmark (1006 human-verified items, 37 models). Shows that fine-tuning on a Cyrillic-script language damages Azerbaijani, while language relatedness predicts nothing | Python · LLM Eval · NLP |
| DSF-Net | Dual-stream spectral-spatial network for detecting AI-generated images (CIFAKE): 95.71% accuracy, 13.2x smaller than ResNet-18, with a two-run replication study | PyTorch · Computer Vision |
| origamicp | Recovering origami crease patterns from images. Mountain/valley is identifiable only up to a global flip; a flip-invariant objective reaches 0.972 up-to-flip accuracy | PyTorch · U-Net · Geometry |
| Apar Feedback Routing | Multilingual AZ/RU/EN support-ticket routing: TF-IDF baseline to XLM-R blend at 0.9186 macro-F1, CPU-only inference under a 600M-parameter budget | Python · XLM-R · ONNX |
| Soccer Tactical Analysis | Voronoi spatial analytics and an explainable-AI tactical analyst powered by the IBM Granite LLM | Python · LLM · Geometry |
| CBAR ETL | ETL pipeline for official AZN exchange rates from the Central Bank of Azerbaijan, with SQL analytics and a Streamlit dashboard | Python · PostgreSQL · Docker |
| Origami Recognition System | Scraping, labelling and dataset pipeline for origami images, the foundation for a fold-recognition model (work in progress) | Python · BeautifulSoup |
| MNIST Digit Recognition (ANN) | Feed-forward neural network built from scratch with manual backpropagation, grid search and 5-fold CV | Python · NumPy |
| Data Science from Scratch | Core ML algorithms (K-Means, GMM, SVM) implemented from first principles | Python · NumPy |
| Competitive Programming | Solutions across Codeforces, CSES and Eolymp | C++ · Algorithms |
- One Bit Per Sheet: Identifiability and Coherence in Crease-Pattern Recovery. Preprint, SSRN, 2026. ssrn.com/abstract=7291678
- AZ-Eval: a parallel Azerbaijani-English benchmark. Dataset and code, Zenodo. doi:10.5281/zenodo.22695091
- origamicp: code and data for crease-pattern recovery, Zenodo. doi:10.5281/zenodo.21848888
- 🥇 7+ awards & honors across competitions and programs
- 💻 ICPC (Azerbaijan): competitor, now coaching the Baku State University teams
- 🤝 HackerRank Campus Crew at Baku State University
- 🤖 Robotex: robotics competition placements
- 🧪 Hackathon participant with hands-on data-analysis experience
| Track | Toolkit | Signature Work |
|---|---|---|
| LLM Evaluation | Python, Transformers, bootstrap & permutation tests | AZ-Eval: measuring how models lose a low-resource language |
| Deep Learning | PyTorch, TensorFlow | DSF-Net, flip-invariant crease-pattern recovery |
| Vision + Geometry | OpenCV, NumPy | Geometric pattern detection and image analysis |
| Data Engineering | PostgreSQL, Docker, Streamlit | End-to-end pipelines and dashboards |
| Math for AI | Graph theory, probability, linear algebra | Bridging theory with practical AI systems |
"Mathematics is not about numbers, equations, or algorithms. It is about understanding." (W. P. Thurston)