Open to work · Belgrade / Remote / UAE

Uroš Milutinović

AI & Software Engineer

Belgrade, Serbia
View Projects GitHub Email Download CV

About

I'm an AI & Software Engineer with hands-on experience building production-grade RAG systems, LLM-powered pipelines, and full-stack web applications. I like working across domains — from backend APIs and mobile integrations to AI model fine-tuning.

Based in Belgrade, open to remote roles and relocation to UAE.

Experience

AI Junior Engineer · Netconomy, Belgrade

Mar – Aug 2026

Built, tested, and refined AI/LLM-based applications and workflows as part of a production-oriented engineering team. Collaborated cross-functionally to integrate AI technologies into the company's software systems within an Agile process.

AI & Full-Stack Engineer · Self-employed / Remote

Dec 2024 – Present

Designed and shipped a RAG system for document Q&A (LangChain, GPT-4, Pinecone) running at 500+ queries/day, with a FastAPI backend and a React.js streaming frontend. Fine-tuned open-source LLMs (Mistral 7B, Llama 3) with LoRA, evaluated via RAGAS/BERTScore. Delivered live client websites and ran marketing campaigns alongside B.Sc. studies.

Projects

🧪

RAG Eval Lab RAG evaluation · 489/508

Took a deliberately weak RAG system from 6/12 to 12/12 one measured change at a time: spec and frozen test questions committed before the code, regression check after every change. 96% across 22 question sets including blind sets, version-aware retrieval for historical policy questions and prompt-injection guards. Hybrid search and query rewriting were measured and rejected.

Python Ollama bge-m3 LLM-as-judge React Playwright
🔧

carservice-api Spring Boot API

Layered Spring Boot backend for a car-service workshop — customers, vehicles, a service catalog, and work orders driven by a status state machine enforced at the entity level, with price-snapshotted line items and invoice generation. 30 tests (JUnit 5, Mockito, full MockMvc integration against a real database).

Java 17 Spring Boot 3 JPA/Hibernate MySQL Flyway Docker
🛡️

MailGuard Security CLI · Active

Flags whether inbox emails are safe to open: layered heuristics over full headers — SPF/DKIM/DMARC, brand-spoofing and lookalike domains, masked links, dangerous attachments — plus an LLM second-pass verdict on borderline cases. Runs as a scheduled daily scan.

Python IMAP SPF/DKIM/DMARC LLM triage launchd
📈

Serbian LLM Benchmark Published · Live

8 open-source models evaluated head-to-head on a Serbian-language task suite — accuracy scored and published as a public leaderboard, with a repeatable one-command pipeline to re-run the full evaluation.

Python Ollama LLM Evaluation GitHub Pages
💳

Credit Card Fraud Detection ML + API

End-to-end fraud detection on 284k transactions (0.17% fraud): KNN, XGBoost, and LSTM in a soft-voting ensemble with SMOTEENN resampling — F1 0.82 on held-out data. Deployed as a FastAPI scoring service (p50 3.5ms) plus an interactive Gradio demo.

scikit-learn XGBoost PyTorch FastAPI Gradio
🔎

Semantic Search over 50k ArXiv Papers NLP

Semantic search engine over 50,000 ML paper abstracts: sentence-transformer embeddings with cosine ranking answer conceptual queries sharing zero keywords with the results, plus related-paper recommendations and a t-SNE cluster map of the corpus.

sentence-transformers scikit-learn NumPy
🎯

JobScout AI Agent

Autonomous agent that reads a job posting, extracts structured requirements with Claude, scores the match on a calibrated 0–100 rubric with gap analysis, drafts a tailored cover letter, and tracks every application — including per-application API cost — in SQLite.

Python Claude API Structured Outputs SQLite pytest
🍽️

dine-agent AI Voice Agent

Claude-powered phone host for a restaurant. Every booking write goes through a policy-owning engine — service windows, table capacities, dining durations, conflict detection — so the model can propose but never invent a table.

Python Claude API (tool use) FastAPI Web Speech API pytest
🔌

poslovi-mcp MCP Server

Model Context Protocol server and CLI that lets any AI assistant search Serbian job boards (Infostud, HelloWorld.rs) through structured tools — HTML parsed into typed results, no API keys required.

Python MCP httpx BeautifulSoup CLI
📊

llm-eval Eval Framework

Compares LLMs (Claude, local models via Ollama) on the same task suite — accuracy, cost, and latency side by side. YAML-defined evals, async runner with token/cost tracking, JSON snapshots for regression detection, and CI-gating exit codes.

Python asyncio Claude API Ollama YAML
🩺

Razumem Nalaz PWA · Live

Progressive web app: photograph a medical lab report and get a plain-language explanation of each value. OCR and interpretation run fully on-device — no data ever leaves the phone.

JavaScript PWA On-device OCR GitHub Pages
🚗

AutoZen Diplomski rad

Web aplikacija za premium auto-detailing studio — online zakazivanje termina, admin panel sa CRUD operacijama, zaštita sesijama i responzivni dark-mode dizajn. Razvijeno kao završni rad na VIŠER-u.

PHP MySQL Bootstrap 5 HTML5/CSS3 Apache/XAMPP
🛡️

SafeCircle

Safety app for online dating — real-time location sharing with trusted contacts, one-tap SOS, check-in system, and reverse image search for fake profile detection. Privacy-first: location data is deleted when the session ends.

Python FastAPI SQLAlchemy JWT Twilio React Native
📞

Alarm Callers

Two Python scripts for repeated automated calls — one via macOS FaceTime (no external accounts needed), one via the Twilio API. Built as a reliable wake-up alarm that can't be silenced.

Python Twilio API macOS Automation

Skills

AI / ML

Python PyTorch LangChain RAG Systems LLM Fine-tuning Hugging Face OpenAI API Pinecone

Backend

FastAPI Spring Boot Java PHP PostgreSQL MySQL SQLite

Frontend

React.js React Native JavaScript HTML / CSS

DevOps

Docker Git / GitHub Linux

Contact

Open to AI/Software Engineer roles — remote or relocation to UAE.

Email umilutinovic25@gmail.com
GitHub umilutinovic25-hash