AI Engineer & Data Analyst — every project ships with its numbers.
I build local AI systems, train and evaluate models, and analyse datasets in Python and R. Then I write down what the results actually show: a metric table, an error analysis, or the limitations of the run.
What I work on
-
AI engineering — LLMs, QLoRA fine-tuning, agents, retrieval, computer vision
-
Data science — modelling, evaluation design and error analysis
-
Data analytics — R, SQL, exploratory analysis, statistics and reporting
-
🔭 Currently: evaluating fine-tuned LLMs (QLoRA on Qwen2.5) and dataset quality for computer vision
-
🌍 Based in Azerbaijan · code and reports in English, tutoring content in Azerbaijani
-
💬 Open to collaboration on local-first ML projects
| Project | What it does | Stack | Result |
|---|---|---|---|
| Base vs Fine-Tuned Panel | One prompt, two variants of the same model, side-by-side answers with measured latency | Transformers, PEFT/QLoRA, Streamlit | Adapter: +0.17 s load, +10.4% latency; quality mixed, documented per prompt |
| Negation & Sarcasm Stress Test | 15 hand-written sentences that break sentiment models in specific, explainable ways | Python, scikit-learn, Transformers | TF-IDF 40% failure · BERT 46.7% · transformer fails every sarcasm case |
| Data Collection Quality Ablation | Two YOLO detectors with identical architecture and hyperparameters — only the dataset changes (300 images / 4 sessions vs 800 / 15+ sessions) | Python, YOLOv8, OpenCV | Diversity beat size on a clip from conditions in neither training set |
| ReviewPulse | Sentiment app for small sellers: one review or a CSV batch, with confidence and a batch summary | scikit-learn, Streamlit, pandas | Capstone — classical NLP delivered as a usable product |
| Local Research Agent | Plans, calls tools through Ollama's native tool API, and cites only evidence it actually retrieved | Python, Ollama, JSONL traces | Fake citations invalidate the run; every step is inspectable |
| License Plate Recognition | Detect → crop → OCR pipeline with an honest end-to-end evaluation | YOLOv8, Tesseract, OpenCV | mIoU 0.888 · recall 100% · exact match 15% (3/20) |
Qwen Python Tutor — local Azerbaijani Python tutor built on a 4-bit QLoRA adapter over Qwen2.5-1.5B-Instruct.
Transformers PEFT QLoRA Streamlit — 72-example dataset, r=4, adapter 2.09 MB
Retrieval-Augmented Generation — from-scratch RAG over a real AzTU examination-regulations PDF, no framework.
multilingual-e5-small FAISS Cross-encoder reranking Ollama
Product Review Sentiment — classical TF-IDF baseline vs a pretrained transformer on the same saved Yelp split.
scikit-learn BERT pandas — identical reviews, labels and test set for both
AI Ticket Bot — Telegram support desk where a local LLM classifies tickets as JSON, with a confidence gate and a logged fallback path.
n8n Ollama Google Sheets — see also the rule-based version it replaced
FAQ Chatbot — TF-IDF + cosine similarity over a 12-question FAQ, with stemming and a fallback threshold.
NLTK scikit-learn Streamlit — CodeAlpha AI internship, Task 2
Language Translation Tool — Streamlit translator with input, size and language validation.
Streamlit REST API — CodeAlpha AI internship, Task 1
Pen/Pencil Webcam Detector — fine-tuned YOLOv8n for one custom class, live webcam overlay with FPS.
Python Ultralytics YOLOv8 OpenCV
CNNs & Image Classification — a CNN trained from scratch against an MLP, same seed and schedule, plus five hand-drawn digits as a reality check.
PyTorch CNN MNIST — CNN 98.73% vs MLP 97.36%
Object Detection & Tracking — webcam/video input, YOLOv8n detections, ByteTrack IDs.
Python YOLOv8 ByteTrack OpenCV — CodeAlpha AI internship, Task 4
Music Generation — LSTM/Transformer MIDI generation with audio rendering.
PyTorch MIDI SoundFont — CodeAlpha AI internship
Actuarial Claims Reserving — Chain Ladder reserving with loss development factors and a tail factor, written up with formulas.
Python pandas SQL — estimated unpaid reserve 392,974.60 AZN (synthetic data)
Bank Marketing Campaign Analysis — chi-square testing and logistic regression on a bank's campaign data.
R Logistic Regression R Markdown
Retail Sales Analysis — cleaning, EDA, anomaly detection and visualisation of Azerbaijani retail data.
R dplyr ggplot2 — Code Academy Data Analytics
EDA & Data Visualization — reusable exploratory workflow: missing values, outliers, correlations, hypotheses.
pandas NumPy Matplotlib Seaborn SciPy
- Fix the comparison. Same data, same seed, same hyperparameters — change one thing.
- Measure the failure. A metric table beats an adjective; error analysis beats a demo.
- Write the limitations. Every README has a section about what the project does not prove.
- Keep it local. Most projects run offline on a laptop, with no paid API in the loop.
- CodeAlpha AI internship — four tasks: translation tool, FAQ chatbot, object detection & tracking, music generation
- Code Academy — Data Analytics — R retail-sales analysis and reporting
- Self-directed projects — weekly builds in NLP, computer vision, RAG and agentic systems, each closed with an evaluation report