I'm a data scientist and AI engineer interested in AI research.
Data Science and AI specialist with hands-on experience fine-tuning Transformer models across two industrial research internships — handwritten text recognition for Azerbaijani, and a customer-support chatbot. Works in PyTorch, Hugging Face Transformers, TensorFlow, Pandas and NumPy, on top of strong Oracle SQL and PL/SQL foundations.
Double master's degree in Data Science and Artificial Intelligence. Also builds and ships full-stack applications in React and Node, which helps in delivering models as usable products.
An end-to-end OCR pipeline for handwritten Azerbaijani, a low-resource language with no usable existing system. Two-stage fine-tuning of TrOCR reached 3.47% character error rate, against 17.23% for a CNN–BiLSTM–CTC baseline.
Fine-tuning DialoGPT-small into a support assistant under a hard 4 GB VRAM ceiling — 20,000 prompt–response pairs mined from a 3M-tweet corpus, with the model's limitations documented rather than hidden.
Built an end-to-end handwritten text recognition (OCR) pipeline for Azerbaijani, a low-resource language, to automate digitisation of handwritten real-estate documents.
Fine-tuned DialoGPT-small, a 117M-parameter GPT-2 variant, into a customer-support chatbot using Hugging Face Transformers and the Trainer API.
Developed and fine-tuned NLP models for text classification, sentiment analysis and sequence processing using TensorFlow / Keras.
Two-stage fine-tuning took character error rate from 17.23% to 3.47% on a language with almost no training data. The interesting part is not the architecture — it is that 5,000 real lines were worth as much as twenty thousand synthetic ones, and that most of the error left over sits on seven letters.