Illia Yermachenkov
Offer Illia work on your next project.
Private message
Offer work
Krivoi Rog,
Ukraine
1 day 11 hours back
available for hire
on the service 1 year
Rating
Successful projects
Average rating
Skills and abilities
Portfolio
-
Hybrid Movie Search System
AI & Machine Learning
Developed a hybrid multimodal movie search system that retrieves films using images (screenshots, frames), text descriptions, or combined queries. Built a large-scale pipeline with extensive data cleaning and deduplication, resulting in ~8,000 high-quality movies and 40,000+ vector embeddings. Used CLIP for joint image-text representations and FAISS for fast similarity search. Supports image-only, text-only, and weighted image+text retrieval with high accuracy.
#machinelearining #searchengine #computervision #python
-
Mental Health Assessment System
AI & Machine Learning
About the Project
A machine learning–powered web application designed to estimate the likelihood of depression based on user-provided lifestyle and demographic factors.
The system analyzes inputs such as demographic information, academic workload, sleep habits, nutrition patterns, and more to generate an instant prediction along with general recommendations.
… Tech Stack
Machine Learning & Data Science
• Python, Pandas, NumPy, Matplotlib, Scikit-learn: data cleaning, preprocessing, visualization, model training and comparison
• Pipeline: complete data-processing workflow (missing-value imputation, feature scaling)
Backend
• Flask (Python): REST API for communication with the ML model
• Integration: serves the production React build directly from Flask (Single-Server Deployment)
Frontend
• React + Vite + TypeScript: fast, modern, strongly-typed client interface
• Tailwind CSS: component styling and responsive design (mobile-first)
• React Select: customized UI components for improved usability on mobile devices
Project Highlights
Achieved strong F1-score/Accuracy through careful feature engineering and validation.
Developed a clean, user-friendly form with real-time validation and dynamic result visualization.
The system outputs not only a binary prediction but also the model’s confidence percentage.
Integrated a list of psychological support hotlines for user awareness and safety.
Full mobile compatibility, including fixes for native select issues on iOS/Android.
#machinelearining #ML #python #flask #React/TypeScript #React #TaillwindCSS
-
Automatic License Plate Recognition System (ALPR)
AI & Machine Learning
About the Project
This project focuses on building a computer vision system for automatically identifying vehicles by reading their license plates. The system operates in two stages: first, it detects the plate in the image, and then it recognizes the text on it.
I trained a detection model on a custom dataset, achieving high localization accuracy.
Additionally, I integrated PaddleOCR to extract characters from cropped plate images.
… A Python script was implemented using OpenCV to visualize the results (bounding boxes + recognized text) and to filter predictions based on a confidence threshold.
Inference Pipeline
1. The input image is passed to the detection model.
2. The predicted coordinates (xyxy) are used to crop the region of interest (ROI) containing the license plate.
3. The cropped fragment is processed by PaddleOCR for text recognition.
4. The output is filtered using a confidence threshold (conf_thresh = 0.5).
Visualization
Using OpenCV, bounding boxes and recognized text are drawn on the original image, and the result is saved locally.
Tech Stack
• Language: Python
• ML/DL Frameworks: PyTorch, PaddlePaddle
• CV Libraries: Ultralytics (YOLO), PaddleOCR, OpenCV
#machinelearining #computervision #ML #AI
-
Human Head Detection System
AI & Machine Learning
About the Project
This project focuses on the tasks of crowd counting and safety monitoring through accurate detection of human heads. This is a fundamental component of modern video surveillance systems, where it is crucial to determine the number of people even in dense crowds.
Training Pipeline
… • Model: YOLOv11n (the fastest version, optimized for edge devices)
• Epochs: 100
• Device: GPU (CUDA)
• Result: Consistently high accuracy (mAP50 ≈ 0.97)
Visualization
A custom OpenCV-based function renders bounding boxes and displays the model’s confidence score above each detected object.
Training Results
The model demonstrates strong generalization capabilities. Loss function graphs show stable convergence.
The mAP50 (Mean Average Precision @ 50% IoU) score exceeds 0.9, which is considered an excellent result.
The mAP50-95 metric reaches 0.65+, which is a strong performance for lightweight detectors.
Tech Stack
• Core: Python
• ML: Ultralytics YOLOv11, PyTorch
• CV: OpenCV
#machinelearining #computervision #ML #AI
-
Recommendation systems for restaurants
AI & Machine Learning
The work is focused on developing a recommendation system for dining establishments (restaurants) for individual users. The main focus of the project was on preprocessing a complex, multi-table dataset and building a model capable of predicting whether a particular user would like a specific restaurant.
The goal was to create a prediction function (recommendation) of a restaurant rating for a user based on user profile data and their previous ratings. In other words, performing a User-based collaborative filtering task using KNN.
… Steps performed:
• Loading and merging data from five different sources.
• Detecting and handling missing and inconsistent values in the data.
• Cleaning and normalizing geographic and meta-data of users and restaurants.
• Building the recommendation system model.
Data used
The project uses the dataset restaurant-data-with-consumer-ratings, which consists of 9 tables, out of which five interrelated CSV files were used:
1. geoplaces2.csv: Geographic characteristics and attributes of establishments (e.g., price range, alcohol, dress code, etc.).
2. rating_final.csv: Final user ratings of establishments (overall rating, food rating, service rating).
3. userprofile.csv: Demographic and behavioral profiles of users (e.g., alcohol consumption, nicotine usage, budget, etc.).
4. usercuisine.csv: User cuisine preferences.
5. chefmozcuisine.csv: Types of cuisine offered by establishments.
Modeling
• Applied the concept of User-Based Collaborative Filtering (UBCF), where KNN was used to find the K-nearest neighbors based on cosine similarity metrics.
• Optimization: Addressed scalability issues of classical UBCF using optimized methods; confirmed the need to switch to Weighted Average for accurate rating prediction based on neighbor distances and indices.
• Recommendation function: Developed the function recommend_restoran(), which takes a target establishment (target_placeID), a target user (target_userID), and the number of neighbors (k) to predict the rating.
• Recommendation criterion: The predicted rating is rounded, and if it equals or exceeds a threshold value (recommend_threshold), the restaurant is recommended.
Technologies used
Programming language: Python
Libraries:
• pandas, numpy: Data processing and manipulation.
• requests: For calling an external geo-service (GPS-Coordinates.net).
• sklearn (Scikit-learn): For similarity metrics.
P.S. The screenshot shows the number of missing values in the tables I worked with, the final calculation matrix, and 3 results of the function’s execution.
#ML #machinelearning #datascience #database #analytics #dataanalysis
-
Currency Converter with Real-Time Exchange Rates Update
Web Programming
I developed a full-featured currency converter website that provides real-time exchange rates. This project demonstrates my skills in creating interactive web applications with a comprehensive backend and a reliable database.
Intuitive Frontend with React.js:
- Developed a responsive and user-friendly interface using React.js, ensuring smooth interaction and an attractive appearance.
… - Implemented adaptive layout for correct display on various devices (desktop, tablets, mobile phones).
- Introduced interactive elements for convenient currency selection, amount input, and display of conversion results.
- Added site search functionality for quick access to necessary information.
Reliable Backend with Node.js and Express:
- Built a powerful and efficient RESTful API using Node.js and the Express framework. This API acts as a bridge between the frontend application and the database, providing a standardized way of interaction.
- Ensured secure user authorization with three access levels: Administrator, Premium user, free subscription user, and unauthenticated user. This allows flexible management of features and data access depending on the user role.
- Used JSON Web Tokens (JWT) for managing user sessions and secure authentication, ensuring user data protection.
- Implemented HTTP cookies mechanism to maintain authorization state and enhance user experience.
- Developed full CRUD (Create, Read, Update, Delete) functionality for data management, providing full control over application information (e.g., user or currency management).
- Implemented Server-Sent Events (SSE) for real-time currency rate updates. This allows instant access to current data without constant client requests, which is critical for a converter application.
Data Management with MySQL:
- Used a relational database MySQL for storing and managing user data and currency rates.
- Designed and implemented the database structure, including tables for users (with password hashing and roles) and current currency rates.
-
ATB supermarket chain discount product parser bot
Bot Development
A Telegram bot has been created for automatic monitoring and notifying users about discounts on products of interest from an e-commerce platform.
The bot provides the user with:
- browsing product categories with current discounts;
… - the ability to add and remove selected product codes;
- subscription to automatic notifications about price drops on selected products;
- interaction through inline buttons, commands, and FSM.
Technology stack:
- Python, aiogram (v3)
- BeautifulSoup, requests - for parsing
- dotenv, json, os - for configuration and file handling
- APScheduler - for task scheduling
- asyncio, FSMContext - asynchronous logic and state management
#Telegram #parse #python
Activity
| Latest proposals 10 | Budget | Added | Deadlines | Proposal | |
|---|---|---|---|---|---|
|
Changes to the website
80 USD
|
|||||
|
File download from the website (parsing / data export)
83 USD
|
|||||
|
Restoration / removal of password for file *xlsx
23 USD
|
|||||
|
Parsing and filtering of the skins market in Steam
100 USD
|
|||||
|
Computer vision service
3000 USD
|
|||||
|
Convert Time
500 USD
|
|||||
|
License Plate Recognition
700 USD
|
|||||
|
Parsing Amazon → XML for Prom.ua + Excel
68 USD
|
|||||
|
Website parsing
23 USD
|
|||||
|
Calculator on Windows Forms C#
16 USD
|
