Ірина

KATANA IRYNA
JUNIOR MACHINE LEARNING ENGINEER

CONTACT SUMMARY
Highly motivated Junior ML engineer with a strong
MOBILE: [открыть контакты](см. выше в блоке «контактная информация») foundation in developing machine learning models for
different purposes. Have experience in Prompt Engineering
TELEGRAM: https://t.me/kogarashiiiii
for enhancing the efficiency of different processes.
EMAIL: [открыть контакты](см. выше в блоке «контактная информация») Proficient in Python and C(++), skilled in leveraging libraries
like TensorFlow, Keras, and Pandas to achieve high-quality
LINKEDIN: [открыть контакты](см. выше в блоке «контактная информация») outcomes. Has good communication skills, and is a highly
responsible, flexible, and active individual.
ADDRESS: Kyiv, Remote

EXPERIENCE Feb 2024 -
SKILLS JUNIOR PROMPT ENGINEER | GOIT June 2024

In my role as a Prompt Engineer at GoIT, I was
PROGRAMMING LANGUAGES
responsible for optimizing the homework review process
Python through the development and implementation of AI-
Java driven tools, utilizing the OpenAI API. My primary focus
C++ was on improving efficiency and accuracy in educational
assessments, namely:
DATABASE MANAGEMENT SYSTEMS
Developed automated solutions for checking
MySQL homework assignments, focusing on creating,
SQLite testing, and refining AI-driven prompts.
ArangoDB Analyzed complex homework requirements to
MongoDB identify opportunities for automation that contribute
to more efficient learning processes.
ML/AI-RELATED Worked collaboratively with educators and students
OpenAI API to collect and analyze feedback, which is used to
ChatGPT improve the accuracy and functionality of the
Tensorflow prompts.
Keras Suggested and implemented modifications to
educational content to better accommodate AI-
Pandas
driven automation, facilitating easier and more
NumPy
effective future integrations.
Yolo5
Achievements:
CV2
Developed prompts that enabled the company to
Machine Learning save over 60,000 UAH per month.
Convolutional Neural Optimized prompts, significantly enhancing their
Network Computer Vision accuracy and reducing the need for manual edits to
Natural Language Processing AI responses.
Time Series Forecasting
RNNs
GRUs
EDUCATION
LSTMs BACHELOR OF COMPUTER SOFTWARE ENGINERING
Augmentation
Zaporizhzhya National University
Transfer Learning
2021-present
FRAMEWORKS
Django LANGUAGES
AngularJS
English - Upper-intermediate
OTHER
Ukrainian - Native speaker
Git, JSON, XML, HTML/CSS
PROJECTS

GLASSES SEGMENTATION DARTS DETECTION AND SCORING
PROJECT June 2024 SYSTEM June 2024

Description: I have created a glasses segmentation Description: An advanced computer vision-based
system with high accuracy using advanced deep system was developed for darts game analysis,
learning techniques. The objective was to create a capable of accurately determining scores and
model capable of accurately segmenting identifying winners (red or green) using deep
eyeglasses in facial images. The system utilizes a learning and image processing techniques. The
U-Net architecture, renowned for its effectiveness project encompasses data creation and labeling,
in image segmentation tasks. The original images image preprocessing, object detection with
were preprocessed by resizing them to 256x256 YOLOv5, dartboard detection, and scoring. A
pixels, normalizing, and converting segmentation dataset of 83 images was created, labeled using
masks to one-hot encoding. The model was Label Studio, and preprocessed by resizing and
implemented using the TensorFlow and Keras splitting into training and validation sets. The
libraries, trained with the Adam optimizer and YOLOv5 model was employed for object detection,
sparse_categorical_crossentropy loss function over trained over 200 epochs with a batch size of 16
100 epochs with a batch size of 24. Custom using transfer learning. For score calculation, the
callbacks were used to save the best model and Hough Circle Transform was used to detect the
stop training if there was no improvement after 6 dartboard, and scoring zones were defined based
epochs. The system achieved remarkable accuracy, on the Euclidean distance from the dart to the
with an overall accuracy of 0.99, and IoU and dice dartboard center. The system was evaluated on the
scores for the glasses class at 0.84 and 0.91 test set, producing accurate scores for each dart.
respectively.
Tools&Technologies: Python, YOLOv5, Transfer
Tools&Technologies: Python, TensorFlow, Keras, Learning, Computer Vision, OpenCV, Object
Deep Learning, Image Segmentation, U-Net, Data Detection, Data Labeling, Jupyter Notebook, VS
Preprocessing, VS Code Code

HANDWRITING RECOGNITION SNAKE GAME REINFORCEMENT
SYSTEM Jan 2024 LEARNING Nov 2023

Description: I have successfully developed a highly Description: This project aimed to develop a Snake
accurate handwriting recognition system that is game trainer using reinforcement learning
capable of interpreting handwritten words using techniques powered by artificial intelligence. To
the IAM dataset. The system implements a achieve this, a Snake game environment was
Convolutional Neural Network (CNN) with Long created, followed by an agent that was trained
Short-Term Memory (LSTM) layers and a using a linear Q-learning neural network
Connectionist Temporal Classification (CTC) loss architecture. The neural network consisted of 11
function, which is perfect for text recognition tasks. input neurons, 256 hidden neurons, and 3 output
After thorough training for more than 200 epochs, neurons. The inputs provided essential information
the model was tested on a sample of the validation such as the snake's spatial orientation, trajectory,
data and saved for future use. The system was and proximity to potential hazards and rewards.
implemented using the Python programming During the training process, the agent learned to
language, TensorFlow framework, and custom make decisions on the next move by maximizing its
mltu library, achieving high accuracy rates in text expected cumulative reward. This was achieved
recognition tasks. The project involved meticulous through iteratively playing the game, observing the
data preprocessing, model training, and evaluation, outcomes, and updating the agent's policy to
resulting in a robust system that can accurately improve its performance. The model's performance
recognize a wide range of handwriting styles. steadily increased from 0 to 58 over 100 iterations.

Tools&Technologies: Python, TensorFlow, Keras, Tools&Technologies: Python, Anaconda, PyTorch,
Machine Learning, CNNs, Image Processing, VS Pygame, Machine Learning, Reinforcement
Code Learning, VS Code
E-LIBRARY Sept 2023 SCIJOURNAL MANAGER Apr 2023

Description: The E-library management system is a Description: SciJournal Manager is a desktop
digital solution that efficiently manages library application designed for science journal
resources and users. It is designed to streamline management. It facilitates seamless article
and organize routine operations related to submissions, ensuring each submission includes at
preserving and providing access to library least one author, two abstracts (in English and
resources. This system covers several aspects, such Ukrainian), keywords, and a full-text file stored
as managing the book inventory, automated securely on a server. The system automates the
registration of book issuance and returns, and assignment of two reviewers per article, who
provides a user-friendly interface for interaction evaluate submissions on several criteria, providing
between library staff and readers. ratings and detailed reports for editorial
consideration and author feedback.
Tools&Technologies: Python, JavaScript, Django,
AngularJS, HTML/CSS, SQLite, VS Code Tools&Technologies: Python, ArangoDB, VS
Professional, Arango Shell, Arango Management
(web) Interface

CERTIFICATIONS
DEEPLEARNING.AI TENSORFLOW DEVELOPER

DeepLearning.AI on Coursera
June 2023 - Aug 2023

INTRODUCTION TO TENSORFLOW FOR ARTIFICIAL INTELLIGENCE, MACHINE LEARNING, AND DEEP LEARNING

CONVOLUTIONAL NEURAL NETWORKS IN TENSORFLOW

NATURAL LANGUAGE PROCESSING IN TENSORFLOW

SEQUENCES, TIME SERIES AND PREDICTION