My Projects

FindWay is a globally oriented, asynchronous search and data platform designed to collect, index, and surface time-critical user-generated cases involving missing persons, animals, and stolen property across borders. The system operates as a unified data engine with multilingual, region-agnostic data models and cross-border case handling. Built from scratch with Ruby on Rails, scalable media storage using AWS S3, and an architecture optimized for high-volume writes and real-time discovery. The platform is designed to integrate AI-assisted moderation microservices for text and media filtering, operating as independent, scalable components within the system. FindWay is engineered for reliability, data integrity, and operational resilience in high-stress, real-world scenarios, including crisis and wartime conditions.

A master's thesis project: a fully interactive R/Shiny dashboard tracking vehicle registrations in Ukraine (2017–2023). Features include dynamic plots, time series forecasting, linear regression, model-(ARIMA), and advanced ETL workflows. Designed for policymakers, economists, and automotive analysts.

The system supports user-generated posts, with full commenting functionality at multiple levels, including comments on posts and discussions within thematic threads. Content is structured around topics, allowing users to create, discuss, and moderate themes as independent discussion units, resulting in complex polymorphic relationships between users, content types, and interactions. The platform includes, multilingual interface support.

This is a live presentation site and gateway for the FindWay search platform. • Internationalization: Supports English, Ukrainian, and Polish via flexible YAML-based i18n. • Modern design: Adaptive, clean Bootstrap 5 UI with animated elements. • Production-ready: Dockerized, reverse-proxied with Nginx, and secured with HTTPS via Let's Encrypt. • Role: Full cycle — from idea and design to infrastructure and deployment.

Module (YOLOv8, DINOv2, LoFTR, EfficientNet-B0, ArcFace, Multimodal Biometrics) The module is a multi-model computer vision pipeline integrating YOLOv8, DINOv2, LoFTR, EfficientNet-B0, and ArcFace for automated content moderation, human facial identification, and multi-species animal biometric verification (dogs, cats, horses). The computer vision system performs: • real-time object detection and spatial localization (YOLOv8); • 384-dimensional vector embedding extraction for animal biometrics (YOLOv8 + DINOv2); • 512-dimensional vector embedding extraction for human face matching (ArcFace); • detection and filtering of explicit materials and violence (EfficientNet-B0); • high-throughput vector similarity search indexed via PostgreSQL pgvector. Aligned regions of interest are converted into dense L2-normalized vector representations (384D for animal morphology, 512D for human facial geometry). This ensures high-precision identity matching across varying angles and lighting conditions under strict zero-retention privacy standards.

Building a unified experience using Flutter. The app consumes Rails REST API, implementing secure authentication, real-time data sync, and offline-first capabilities.

Deployed on Hetzner Cloud via Docker. Observability stack: Prometheus for metrics collection and Grafana for visual telemetry. AI layer: CNN (EfficientNet/FaceNet) and NLP moderators ensuring policy-first content control.

Data Sentinel is a high-reliability forensic engine designed to detect invoice tampering and legal non-compliance. The system combines deep byte-level PDF vector slicing, Error Level Analysis (ELA) for graphic manipulation detection, and automated validation against government registries (Biała Lista MF, IBAN Modulo 97). Built around a zero-retention data policy, it evaluates document stream risks in real-time via native Policy Rules without persistent binary storage.