An Unreal Engine 5.6 synthetic data generation pipeline developed as part of a larger computer-vision project focused on detecting advertisements in real and synthetic images and video. Its role is to generate the synthetic portion of the training dataset while reducing the need for manual annotation.

The Blueprint-driven system randomizes camera positions and field of view, varies lighting, weather, fog, rain, and day/night conditions, captures rendered frames, and automatically exports matching YOLO object-detection bounding boxes and segmentation polygon labels. Basic occlusion filtering and Python validation overlays are used to help verify label quality.

Although it was created for advertisement detection, the pipeline is not limited to ads. By replacing the labeled target actors and classes, the same system can be adapted to generate synthetic training data for other object-detection and segmentation tasks.

Built primarily with Unreal Engine 5.6 Blueprints, alongside a small C++ helper for saving label files and Python tooling for dataset validation.

View source on GitHub

A free audiobook discovery and listening app built around the LibriVox public-domain library. Auralia provides access to 40,000+ titles and combines a full-featured audiobook player with search, saved books, user profiles, social features, listening statistics, and AI-powered recommendations.

Developed as part of a mobile app development team formed at my high school, Auralia was selected as the best app from Bosnia & Herzegovina in the m:tel App Competition, earning our team the opportunity to represent the country at the regional final in Podgorica, Montenegro.

Built with React Native and TypeScript, with Firebase for authentication and user data, LibriVox for audiobook content, and Gemini-powered personalized recommendations through a custom Auralia API.

An AI-driven ATM demand forecasting and cash-management concept developed for the Raiffeisen AI Hackathon. The system uses time-series analysis to predict when individual ATMs are likely to run low on cash and estimate how much money should be replenished, helping reduce unnecessary idle cash while improving ATM availability.

For bank operations, the solution is designed to prioritize ATM refill deliveries and support Cash-in-Transit route optimization using predicted demand, refill urgency, and geographic clustering. For customers, the concept includes ATM availability insights, personalized alerts, and predictions of periods with increased traffic at branches and ATMs.

The proposed machine-learning pipeline uses Python with pandas and scikit-learn for data processing and modeling, with more advanced time-series approaches such as LSTM networks considered for future development. The model is designed to integrate with existing banking systems through APIs while operating primarily on anonymized transactional data.

The broader goal is to move ATM management from a reactive process to a proactive one: reducing operational costs and downtime, minimizing excess idle capital, and improving the customer experience through better resource planning.

Radio HARF is an independent online radio station and web platform built around a custom live-streaming setup. The site provides a browser-based player with live status, now-playing metadata, listener count, recently played history, and supporting pages for the station's wider content.

The streaming backend runs on a VPS using Icecast for audio distribution, Liquidsoap for playlist automation and live DJ input, and Nginx as a reverse proxy. The station can switch from a live broadcast to an automated randomized music playlist, with a fallback chain designed to keep the stream available even when no DJ is connected.

The frontend is built with HTML, CSS, and JavaScript, while Supabase is used for the site's chat functionality. Additional infrastructure includes HTTPS via Let's Encrypt, systemd services for process management, and API-style endpoints for shared playback history and stream information.

The project is still evolving, with continued work on the station website, archive, blog, interface, and overall reliability of the live radio experience.

Visit radio.kerim.ba