Hyvärinen

High performance workout manager

Demo

Custom-Built Full-Stack Architecture

A highly concurrent, dependency-free workout management system built on a custom Java HTTP server, with secure authentication and a dynamic frontend

Custom HTTP server using Java virtual threads for high-throughput, low-latency request handling
Interactive frontend with a dynamic calendar and workout management
JWT-based session authentication with secure HttpOnly SameSite cookies
Rate limiter with a thread-safe sliding window with per-IP tracking
Custom JSON serialization and deserialization
User authentication with PBKDF2 password hashing and per-user salts
Connection-pooled JDBC persistence with transactional DAOs
Admin dashboard for live system metrics
Asynchronous logger with batching and file rotation

Screen capture translation tool

Custom Desktop OCR and Translation Pipeline

Desktop tool for capturing a screen region, extracting multilingual text via OCR, and displaying an English translation in a dynamic overlay, triggered by a global hotkey

Custom-built "Snipping Tool" interface for screen region selection
Integrated multilingual OCR using Tesseract (Tess4J) with support for multiple scripts
Automatic translation via Google Translate HTTP endpoint
Dynamic Swing overlay that adjusts size and position based on screen bounds
Asynchronous translation handling to prevent UI blocking
System tray support for background execution and clean shutdown
Global hotkey capture with JNativeHook

Object recognition model for Dota 2

End-to-End SSD Object Detection Pipeline

Implemented an SSD300 convolutional neural network to detect heroes, creeps, and towers in Dota2 screenshots, including model training, evaluation, and visualization.

Single-shot detection for fast localization and classification
Custom Dota2 dataset with 6 object classes, bounding boxes, and difficulty tags
PyTorch training pipeline with VGG16 backbone, transfer learning, and GPU acceleration
Image preprocessing, tensor transformations, and post-processing with NMS
Evaluation using per-class AP and overall mAP with visualized predictions

Hybrid ViT & CNN Image Classifier

Custom Vision Transformer + CNN Hybrid Models

Implemented transformer-based image classifiers and hybrid CNN/ViT architectures, including token distillation, serial and late fusion approaches, achieving high accuracy on Fashion-MNIST.

Built Vision Transformer (ViT) from scratch with class and distillation tokens
Implemented serial and late fusion hybrid models combining CNN feature extraction with ViT
Hard distillation training for student-teacher knowledge transfer, boosting accuracy
Achieved 90% test accuracy with late fusion hybrid model in only 5 training epochs
Visualized predictions for interpretability, showing correct classification even on ambiguous labels
Custom training pipelines with PyTorch, GPU acceleration, and feature map handling for hybrid input

Milk tracker

Android app with persistent storage and dynamic UI updates

An android app for tracking a baby's milk consumption and diaper changes.

Tab-based interface for switching between milk and diaper tracking
Persistent SQLite storage with custom database helper and DAO logic
Dynamic UI updates showing totals, today's counts, and time since last entry
Fragment-based architecture with callback-driven UI updates
Time calculations and formatting for daily totals and intervals between events

Eetu Hyvärinen

Software Engineer