Projects

B2B SaaS bringing a face-recognition-driven VIP experience to luxury retail. As co-founder and CTO I own the backend and the entire face-recognition pipeline. The MVP recognizes a VIP customer at the door in under 200 ms and instantly pushes their full profile to a store associate's phone — built on a consent-first, GDPR-compliant architecture that stores only mathematical embeddings, never images.

Face-recognition software for weddings and large events: a guest scans a QR code, takes a selfie, and every photo they appear in is delivered to their private gallery in real time while the event is still running, with AI auto-enhancing image quality. Like Sila, it is embedding-only (stores vectors, not faces) for full GDPR compliance — the differentiator versus PhotoDay and Waldo Photos. Recognized at the Science Technology Park Belgrade innovation competition.

Enterprise-grade budget planning and execution system for the Ministry of Finance, built with Saga / Yotta. Worked as part of a large team following Agile methodology with sprint-based development, contributing across the full lifecycle: backend architecture, database design, automated testing, containerization, and deployment. Also delivered sibling enterprise modules on the same stack and standards — asset inventory, warehouse, and vehicle-fleet systems — and helped modernize the company's internal Java training to Java 21 + Spring Boot 4.

AI-powered system for automated CV generation based on candidate metadata and job requirements. Independently developed under expert supervision, currently in production and actively maintained. Used by HR professionals and recruiters for streamlined candidate presentation.

End-to-end ownership of a live recruiting platform used by paying clients (Spring Boot 3.5 / Java 21 backend, Angular frontend). It deeply integrates the Workable ATS, runs scheduled sync "agents" that ingest all candidates and jobs in batches, verifies and enriches LinkedIn data through a multi-vendor strategy with adaptive throttling, and powers semantic skill search via a Neo4j knowledge graph kept alongside MySQL (polyglot persistence). Clients access their own data through a dedicated, API-key-secured public API.

Custom CLI tool that generates DTO classes and REST endpoints from specially formatted markdown specification files. Built entirely with Java Reflection - pure logic, no AI. The tool parses markdown documents produced during the design phase and automatically generates boilerplate code, significantly accelerating development workflow.

Built and maintained a conversational AI chatbot for Bright English, an established Duolingo-style platform for learning English, adapting and extending it to the client's needs. A FastAPI backend exposes both REST and WebSocket endpoints for text and audio streaming, with text-to-speech that supports barge-in (the user can interrupt) and latency reporting, JWT auth, and a layered architecture that explicitly follows SOLID.

Personal virtual machine with rented domain hosting deployed applications and websites. Self-managed infrastructure with SSH terminal access and custom CI/CD pipelines for personal projects. If you see this portfolio - the VM is working!

Frontend application for a vacation rental property. Features responsive design, image gallery, contact integration, and SEO optimization. Deployed with custom domain configuration on personal VM infrastructure.

Full Stack system developed to support teaching and student performance tracking, assisting with grading and academic records management at the Faculty of Civil Engineering.

A standalone desktop Git GUI built with Tauri (Rust + web frontend) to replace the VSCode Git panel and the terminal — including a workflow for committing and pushing from a phone. It ships with full release engineering: a GitHub Actions build matrix for Windows and Linux, automatic version patching, GitHub Releases with latest.json generation, SCP deploy behind Nginx, and one-line installers with SHA256 verification.

A lightweight PDF e-signing tool: upload and preview PDFs, draw a customizable signature, position/drag/resize the overlay on any page, and download the signed document, with automatic temp-file cleanup. A small, focused full-stack app pairing a Go standard-library backend with a Svelte + Vite frontend.

A collection of small, independent, phone-friendly web apps for organizing an event — a DJ song-wishlist tool with admin curation and one-click audio download, and a guest survey (drinks, food, transport/carpool, song wishes) with an admin purchase- and carpool-planning dashboard. Each app is deployed as its own process on a subdomain behind Nginx, with systemd units and per-app deploy scripts.

Solo-developed financial software for a construction company. GeoManager (Spring Boot + Angular) manages financial transactions, reads XML files, and exports reports to Excel via Apache POI — 1000+ commits to beta/production. Its companion GeoReader reads bank statements, handles contracts and a cashbook, generates browser-based reports for non-technical staff, and ships as a runnable JAR with a local server and an in-app bug-reporting flow.

An empirical, statistically-defensible comparison of AWS, Azure, Google Cloud, and Hetzner across three architectural layers — IaaS, CaaS, and FaaS — using one identical FastAPI application deployed to ten targets. The platform underpins an in-progress PhD paper: a fully automated 14-day measurement campaign on a neutral test client, analyzed with non-parametric statistics. It even caught a critical confound before launch (the storage benchmark was timing SDK client setup, not storage) and cut p95 latency by up to 156x.

Master's thesis research implementing identical microservice architectures in Spring Boot, Micronaut, and Quarkus. Comprehensive comparison across performance metrics, documentation quality, community support, and developer experience. Results published as a scientific paper demonstrating microservice application in geodesy domain.

Patented software for Geosoft d.o.o., developed with professors from the Faculty of Civil Engineering in Subotica, that performs 1D and 2D least-squares adjustment of survey measurements — replacing legacy Excel + VBA templates. I reverse-engineered the original VBA algorithm, wrote a full specification, then implemented a Python (NumPy) core behind FastAPI, proving correctness with golden-file tests against real solved projects (17/17 passing, including an 85-point / 301-measurement dataset, to 6-decimal tolerance).

A pipeline that converts scanned cadastral/geodetic PDFs — largely handwritten Serbian Cyrillic — into clean CSV using two independent extractors (Claude and Google Gemini) from different model families to avoid correlated errors. The results are compared cell-by-cell; disagreements are routed to a human-review UI, and document-internal checksums (group sums, JMBG control digit) act as model-independent validators.

Completed intensive Spring Boot internship program at SAGA, focusing on backend development fundamentals, enterprise patterns, and professional software development practices.