Ognjen Aleksandar Jevtić

Software Engineering student
AI & agentic systems

I build software where language models do real work: AI coaches, retrieval systems, and the full stack around them.

RAF · Belgrade
44.81°N 20.46°E

Scroll

02 About

I'm a Software Engineering student at RAF in Belgrade, working my way toward AI and agentic engineering. I like owning a thing end to end: the model call, the API behind it, and the screen someone actually touches. So far that means AI coaches, a retrieval system over Serbian traffic law, and an exam platform for programming courses.

Studying
Software Engineering
School of Computing (RAF)
Focus
AI & agentic
engineering
Languages
05Python, Java, C,
C#, JavaScript
Status
Open to internships
and junior roles

03 Selected work

Things I've built

01 / 052026

Fitness AI

A chat-first fitness tracker with a coach that knows your data.

Log food, workouts and body weight by typing a plain message, the way you would text a friend. The app works out the calories and protein, keeps your dashboard up to date, and an AI coach answers your questions using your own history. Works on phone and desktop with the same data.

  • React
  • Vite
  • Express
  • Supabase
  • PWA
  • LLM APIs
Fig. 01 · Message → parse → dashboard
1840 KCAL bench 80 4x8 → workout · 4 sets 300g chicken, yogurt → meal · protein logged
02 / 052026

Weak Keys

A typing trainer that finds what you type badly and drills exactly that.

A browser typing test that tracks the keys, words and letter pairs you miss, then builds practice from them. An optional AI coach, Claude or Gemini, reads your whole history and says what is improving, what is stuck and what to drill next. No account, no server, nothing to install.

  • JavaScript
  • HTML
  • CSS
  • Claude API
  • Gemini API
Fig. 02 · Missed keys become the next drill
the quick brwon fox jumps
03 / 052026

Traffic Law RAG

Question answering over Serbia's road safety law.

Three retrieval pipelines compared on the same 362 articles: fixed-size chunks with exact search, article-aware chunks on an HNSW index, and a third that adds cross-encoder reranking and pulls whole articles back hierarchically. Embeddings and reranking run locally; the answer comes from an LLM on Groq.

  • Python
  • FAISS
  • sentence-transformers
  • Reranking
  • Groq
Fig. 03 · Law → vectors → nearest articles
362 ARTICLES FAISS · HNSW ČLAN N
04 / 052026

RAF LMS

An exam platform for programming courses.

A teacher sets a task, students solve it in a desktop editor and hand it in, and the teacher follows the session live and grades with comments on individual lines of code. Three parts: a Spring Boot REST API with JWT and MySQL, an Electron editor built on Monaco, and a React dashboard for monitoring.

  • Java
  • Spring Boot
  • MySQL
  • JWT
  • Electron
  • React
Fig. 04 · Solve, submit, review by line
Zadatak1.java check the bound LIVE SUBMITTED
05 / 052026

Parallel Pareto

One multi-objective search, written three ways.

The same Pareto optimization implemented sequentially, with threads over shared state behind locks, and as separate processes that trade candidates over a ring and other topologies. Each version is measured against the others, with hypervolume as the yardstick for the quality of the front.

  • Python
  • NumPy
  • Threading
  • Multiprocessing
Fig. 05 · The non-dominated front
NON-DOMINATED OBJECTIVE 1 OBJECTIVE 2

04 Stack

Languages

  • Python
  • Java
  • C
  • C#
  • JavaScript

AI & data

  • LLM APIs: Claude, Gemini, Groq
  • RAG pipelines
  • FAISS
  • sentence-transformers
  • Reranking

Backend

  • Spring Boot
  • Node.js, Express
  • REST, JWT
  • MySQL
  • Supabase

Frontend

  • React
  • Vite
  • Electron
  • PWA
  • HTML, CSS

Tools

  • Linux
  • Git
  • CI/CD
  • Threads & processes

05 Contact

Hiring for an internship or a junior role, or building something with AI? I'd like to hear about it.