Sergej Čipak

Desktop app for managing Serbian fiscal invoices. Scrapes purs.gov.rs, decodes QR codes, exports to Excel, and generates receipt images, from a simple GUI. Desktop aplikacija za menadžment fiskalnih računa. Scrapuje podatke sa purs.gov.rs, dekodira QR kodove, exportuje u Excel, i generiše slike računa, sa prostim GUI

The main challenge was reliable QR decoding on Windows ARM (Parallels), pyzbar failed due to missing DLLs, so I replaced it with zxing-cpp and packaged everything with PyInstaller. CI/CD via GitHub Actions builds Windows and macOS releases automatically on every tag push. Glavni izazov je bilo pouzdano QR dekodiranje na Windows ARM (Parallels), pyzbar nije radio zbog nedostajucih DLL biblioteka, zamenio sa zxing-cpp i spakovao sve sa PyInstaller. CI/CD preko GitHub Actions pravi Windows i macOS release automatski na svakom tag push.

Self-hosted Linux server running Home Assistant, Nginx reverse proxy, and containerised services via Docker Compose, managed remotely over SSH. Linux server koji pokrece Home Assistant, Nginx reverse proxy, i docker usluge preko Docker Compose, upravlja se isključivo preko SSH.

Set up from bare metal: partitioning, network config, SSH hardening, firewall rules, Let's Encrypt TLS, and automated Docker and systemd service restarts. Learned the hard way that DNS resolution inside containers needs complex configuration. Konfigursan od nule: partioning, network config, SSH hardening, firewall pravila, Let's Encrypt TLS, i automatski restartovi Docker i systemd servisi. Naučio sam na tezi način da DNS resolution unutar kontejnera zahteva kompleksnu konfiguraciju.

4th place at EESTech Challenge 2026. Built a hybrid autonomous driving agent for the MetaDrive simulator — behavioural cloning neural network for steering, rule-based controller for throttle, and a YOLO + Kalman perception pipeline for obstacle avoidance. 4. mesto na EESTech Challenge 2026. Razvio hibridni autonomni agent za MetaDrive simulator — neuronska mreža trenirana bihejvioralnim kloningom za upravljanje, rule-based kontroler za gas, i YOLO + Kalman pipeline za detekciju i izbegavanje prepreka.

Data collection runs in multiple modes (human, DAgger, human-assist) with per-source sample weighting — human steering corrections weighted 3× higher. The perception module runs adaptively: idle at 1 frame per 60, switching to 1 per 5 when lidar detects nearby objects. Risk scoring uses IoU overlap between predicted object boxes and a projected ego corridor. Prikupljanje podataka radi u više modova (human, DAgger, human-assist) sa ponderisanjem uzoraka po izvoru — ljudske korekcije upravljanja imaju 3× veći težinski koeficijent. Perception modul radi adaptivno: idle na 1 frame od 60, prelazi na 1 od 5 kada lidar detektuje bliske objekte. Procena rizika koristi IoU preklapanje između predviđenih bounding boxova i projektovanog ego koridora.