Full-stack engineer

Projects

01 · Season 1 live

Carsnap

An AR car-spotting game. Point your camera at a car, AI identifies it, and it becomes a collectible battle card with real-world stats. Build a deck and climb ranked seasons in turn-based duels.

• Flutter
• FastAPI
• GCP
• YOLO
• AI

Visit CarSnap ↗

How it's built+

car models

1,000+

AR detection

Real-time

live now

Season 1

• A YOLO-powered camera detects cars in real time, and AI identifies the make and model.
• Every capture becomes a card with real stats: horsepower, handling, weight and more.
• Turn-based ranked duels with modifier cards and a synergy system.
• Matchmaking queue, ranked seasons, daily challenges and achievements.
• Pack opening, a deck builder and a multi-currency economy.

Under the hood

A Flutter app backed by FastAPI services on Google Cloud Platform. A YOLO model detects the car in the camera feed, AI identifies the make and model, and the backend turns it into a card and runs matchmaking, ranked battles and the economy.

• 
• 
• 
• 

02 · Live at sterza.app

Sterza

A driving companion for car enthusiasts: discover and share great roads, plan scenic routes, time circuit laps and get a corner-by-corner debrief after the drive.

• React Native
• Maps & GPS
• AI coaching

Visit sterza.app ↗

How it's built+

GPS sampling

5 Hz

lap time check

±1.0 s

off-track lap rule

40 m

• Road Atlas: trace, save and share driving roads on the web, no account needed.
• Route builder for scenic drives, shared exactly the way they were driven.
• Built-in circuit lap timer with leaderboards; laps more than 40 m off the track do not count.
• Car meets and group drives: a pin, a time and who is coming, with the exact address revealed two hours before.
• Sterza Plus: braking and throttle zones on the map, apex cues, ghost laps and an AI summary of what to change next time.

Under the hood

A React Native app with map-based road discovery, GPS recording at 5 Hz for laps and runs, and an AI coaching summary after each drive.

• 

03 · Play on itch.io

Lacuna

An online trading card game with 442 cards across eight colours. A TypeScript rules engine runs on an authoritative game server, and the Unity client only renders the state and sends back moves the server has validated.

• Unity
• TypeScript
• Socket.io
• Supabase
• Stripe

Play on itch.io ↗

How it's built+

cards

442

colours

8

test files

114

• Server-authoritative: every state update ships the legal actions, so no game rules live in the client.
• Rules engine with a response chain, instant-speed windows, formations, triggers and mulligans.
• Online play through private rooms with codes, plus a practice bot and a turn timer.
• Accounts and collections on Supabase; booster packs, crafting, quests, progression and cosmetics.
• In-game shop with Stripe checkout and webhook-based fulfilment.

Under the hood

A pnpm monorepo: a pure TypeScript engine package, a shared package of types and legal-move hints, a Socket.io game server deployed on Fly.io, and a Unity 6 (URP) client that is a presentation layer only.

• 
• 

04 · Open source

Raft in FastAPI

A Raft-replicated key-value store: four FastAPI nodes elect a leader, replicate every write through an append-entries log, and elect a new leader when the old one goes down.

• Python
• FastAPI
• Distributed systems

View code ↗

How it's built+

FastAPI nodes

4

randomised election timeout

10–20 s

built

2024

• Each node is a FastAPI app exposing the key-value API (GET, POST, PUT, DELETE) and the Raft RPCs.
• Leader election over a /vote endpoint with randomised timeouts and a majority that discounts dead nodes.
• Writes go through the leader and are replicated to followers via /append_entries.
• A heartbeat monitor thread keeps followers in line; followers start an election when heartbeats stop.
• Reads can be redirected to the least-loaded follower.

Under the hood

Run one uvicorn process per node; each node's /docs page exposes every endpoint for demonstrating elections and replication.

05 · Open source

Portal

A Flutter app for running language models on your own machine: browse, install and hot-swap models, chat with streamed responses, tune generation settings per model and see what your hardware can handle.

• Flutter
• Dart
• FastAPI
• Local LLMs

View code ↗

Install model↓

Hot-swap⇄

Save session✓

How it's built+

token-by-token chat

Streaming

generation settings

Per model

hardware check

CPU · GPU

• Model manager: install, activate (hot-swap) and delete models from a catalogue.
• Chat that streams tokens as they are generated.
• Per-model generation settings: temperature, top-p, max new tokens, repetition and length penalties.
• Device panel showing CPU, RAM, GPU and CUDA availability.
• A 3D model viewer and animated video background in the UI.

Under the hood

The Flutter client talks to a local FastAPI backend on 127.0.0.1 that manages the models and runs inference; this repository is the Flutter frontend.

Also built

• SpellchainThe game from those mobile ads, turned into a roguelike endless runner.Unity · C# · Supabase+

  A portrait mobile crowd runner: steer a coven of wizards through gates that grow your army, crack chests for spell powers and outlast a guardian every level. Levels are generated from their number, so level 37 is the same for everyone.

  • Your army is both your health and your firepower.
  • Gates you can shoot to raise their value, chests with spell powers, and an elite guardian every fifth level offering a power or a passive.
  • Seven enemy types, each demanding a different response: focus fire, dodge or kill first.
  • Per-level difficulty curves for enemy health, speed, attack frequency and upgrade costs.
  • Supabase backend with server-side purchase verification.

  • 
• SUI OmnimarginA margin trading plugin for the Sui network.Sui
• FIPU QuizUniversity exam app with AI proctoring that catches photo and video spoofing.Vue · FastAPI · AI+

  A test-taking web app for university exams that uses AI to check the student is present on camera, and detects a photo or pre-recorded video used as the camera feed, or a camera that is too dark.

Plus a handful of private projects in progress. Happy to talk about them.

How I use AI

1. 01

   I own the architecture

   Data models, system boundaries and trade-offs are decided by me before any code exists.

2. 02

   AI multiplies the build

   I use AI agents to move faster on implementation, so I can take on projects bigger than one developer usually should.

3. 03

   Nothing ships unread

   Every change is reviewed and tested by me. I can explain any line in production.

Experience

1. Now

   Software Engineer · Cenosco

   Built the framework for the new AI system that drives the company's legacy application, now the foundation for every new feature.

   • Implemented the framework for the new AI system that controls the legacy application; all new features build on it.
   • Built mission-critical components.
2. 2024

   Intern · Asseco Group

   Java developer intern: a Spring Boot backend service, a Svelte web app and a Kotlin Android app.

3. 2022

   Team Lead · Huawei Seeds for the Future

   Led one of two Croatian teams designing a sustainable energy-saving solution. Selected as the top-performing candidate in Croatia and invited to the Huawei Talent Summit in Athens.

Stack

Languages

PythonTypeScriptJavaScriptDartC#JavaKotlinSQL

Backend

FastAPIDjangoFlaskSpring BootNode.js.NET

Frontend & mobile

FlutterReactReact NativeVueSvelte

Games

Unity

Data & cloud

PostgreSQLMongoDBFirebaseSupabaseGoogle CloudDocker

AI

Local LLMsAI features in productsAI-assisted development

Say hi