CSONGOR OSZ · SOFTWARE ENGINEER

Building AI.
From retrieval
to vision.

I’m Csongor, a Computer Engineering graduate and AI & Vision master’s student in Cluj-Napoca. I build Python services, LLM agents, and tools for evaluating machine-learning systems.

Explore projects View CV

PROJECT SPOTLIGHT

VisionOps Analytics: the project's real-model Grafana monitoring dashboard.

VisionOps AnalyticsComputer vision, measured.

220evaluation images

10food classes

98.2%recorded accuracy

Pretrained classifier · results on this evaluation set

CLUJ-NAPOCA, ROMANIAPython · Applied AI · Full-stack

LLM APPLICATIONS / COMPUTER VISION / BACKEND SYSTEMSSELECTED WORK

01 — SELECTED WORK

Built. Tested. Explained.

Four projects exploring how AI applications retrieve, reason, interact, and get evaluated.

01 / LLM APPLICATION DEPLOYED

Tutor Chatbot.

Help the student find the next step.

A Socratic tutor for programming concepts. It starts with questions and analogies, then escalates through hints before revealing an explained answer.

WHAT I BUILT

A FastAPI and Next.js application combining a LangGraph tutoring workflow, PDF retrieval, streamed responses, and Judge0 code execution.

ENGINEERING DECISION

Track the topic, understanding, and misconceptions in explicit state. Choose the next teaching strategy from four hint levels instead of relying on a single prompt.

IN PRACTICE

The example conversation moves from a recursion analogy to executing a factorial program. Uploaded study material supplies additional context through ChromaDB retrieval.

PythonFastAPILangGraphChromaDBNext.jsJudge0

Try the tutor Source code

Tutoring workflow

Extract topic → assess understanding → choose strategy → respond. PDF retrieval and code execution provide context when relevant.

Hint levels: analogy → targeted hint → leading question → explained answer.

02 / COMPUTER VISION & MONITORING

VisionOps Analytics.

What happens after a model predicts?

An evaluation and monitoring pipeline connecting a pretrained food classifier to PostgreSQL, Jupyter analysis, and Grafana dashboards.

220real images

98.2%accuracy on this set

210/220accepted at ≥ 0.70

WHAT I BUILT

Integrated pretrained Vision Transformer inference with prediction storage, error analysis, confidence thresholds, and Docker Compose provisioning.

ENGINEERING DECISION

Store real and synthetic predictions separately. Use confidence thresholds to explore the trade-off between automatic acceptance and human review.

WHAT THE EVALUATION SHOWED

All 210 accepted predictions were correct on this set. A separate 50,000-record synthetic dataset supports monitoring across simulated versions and locations.

PyTorchHugging FacePostgreSQLGrafanaDocker

Evaluation: 200 Food-101 validation images + 20 manually collected images, across 10 classes. These are results from a small, domain-specific evaluation of a pretrained model, not a production-performance claim.

Explore the project

03 / CONVERSATIONAL AI

Mini AI Interviewer.

A conversation that follows your answers.

A focused, four-question interview that adapts to the participant, optionally uses their CV, and produces a summary with the complete transcript.

WHAT I BUILT

A Next.js interface and FastAPI backend with streaming responses, CV retrieval, JSON reports, and user-requested email delivery through Resend.

ENGINEERING DECISION

Keep CV upload optional and temporary. Validate requests, limit uploads, and expire document sessions while preserving completed reports separately.

THE COMPLETE FLOW

The screenshots follow CV-informed questions through a completed summary to the received email and attached JSON transcript.

FastAPINext.jsGroqChromaDBFastEmbedResend

Try an interview Source code

04 / MCP & TOOL CALLING

Research Paper MCP Agent

A command-line research assistant that discovers arXiv papers, saves metadata and abstracts by topic, and retrieves previously found papers through LLM-selected tools.

Groq LLMMCP client / serverarXiv + JSON storage

Typed tool schemas and asynchronous stdio communication separate the assistant from its research capabilities. Its current scope is metadata and abstracts; full-paper PDF retrieval is future work.

PythonFastMCPGroqarXiv API

Explore source code

Earlier explorations 09

CodeReviewAgent

Multi-step code analysis and structured review reports

FastAPI · LangGraph

Dispatcher Manager

Shipment details, costs, and profit tracking

Python · Django

Social Network

Authentication, posts, follows & personalized feeds

Django · JavaScript

Stock Trading Web App

Simulated trading with live price data

Flask · SQLite

Auctions

Listings, bidding logic & watchlists

Django · SQLite

Bomberman

Game loops, collisions & grid-based logic

C++ · OOP

Library Management System

Books, users & borrowing workflows

C# · .NET · SQL

Mail

An asynchronous, single-page email client

Django · JavaScript

Infrastructure Cost Estimator

Configurable pricing & budgeting logic

Python · CLI

02 — BACKGROUND

From sensors
to intelligent systems.

Ősz György Csongor · Cluj-Napoca

View my CV

My engineering foundations started with hardware as well as software. For my bachelor’s thesis, I built an Arduino irrigation system with soil-moisture sensing, automatic pump control, and a Bluetooth-connected Android app.

That interest in connected systems now shapes my work with AI: how models retrieve information, use tools, retain state, and turn predictions into something people can use.

M.SC. · 2026–2028 EXPECTED

Artificial Intelligence and Vision
Technical University of Cluj-Napoca

Dissertation proposal: multimodal RAG for enterprise knowledge retrieval, combining vision-language models, cross-lingual search, and evaluation.

B.SC. · 2020–2024

Computer Engineering · University of Oradea
240 ECTS · Thesis grade: 9/10

03 — EXPERIENCE

Beyond personal projects.

Software, infrastructure, and the day-to-day responsibility of keeping operations moving.

Plexus

Software Engineer Intern

JUL–SEP 2022

Built a C#/.NET application and scripts to automate incident, task, and back-office processing. Tuned database applications to improve processing performance.

iNet Corporation

IT Engineer Intern

JUL–SEP 2023

Supported LAN/WAN networking, firewall and network security troubleshooting, and server and desktop systems. Assisted with secure wireless configuration.

Tonirad Trans SRL

Shipment Planner

NOV 2024–FEB 2026

Coordinated international shipments and multi-leg routes, tracked exceptions, and resolved disruptions. Experience managing real operational constraints and communicating across teams.

CONTINUED LEARNING

· Hugging Face: Agents Course
· DeepLearning.AI / Anthropic: MCP course
· Harvard CS50: CS50x (Computer Science), CS50P (Python), CS50AI (Artificial Intelligence with Python)

Additional coursework in web programming (CS50w: Web Programming with Python and JavaScript) and SQL (CS50SQL: Databases).
Languages I speak: Hungarian and Romanian (native), English (fluent), German (A2).

04 — MY TOOLBOX

Tools, put to work.

The technologies behind the projects above, grounded in a broader engineering foundation.

Agents & retrieval

Tutor · Interviewer · MCP research

PythonFastAPILangGraphLangChainChromaDBFastMCPGroqGemini

Vision & evaluation

VisionOps Analytics

PyTorchHugging Face TransformersPandasScikit-learnJupyterMatplotlibGrafanaPostgreSQL

Applications & delivery

From interface to deployment

Next.jsReactTypeScriptTailwind CSSDockerGitHub ActionsRailwayVercelDjangoFlask

Additional foundations: C++, C#/.NET, Java, SQL, embedded systems, REST APIs, state machines, and streaming.

05 — GET IN TOUCH

Have something
in mind?

A project, an opportunity, or a good conversation about what’s next. I’d like to hear from you.

csongorosz@gmail.com

LET’S BUILD SOMETHING USEFUL.

GitHub LinkedIn

Building AI.From retrievalto vision.

Built. Tested. Explained.

Tutor Chatbot.

VisionOps Analytics.

Mini AI Interviewer.

Research Paper MCP Agent

CodeReviewAgent

Dispatcher Manager

Social Network

Stock Trading Web App

Auctions

Bomberman

Library Management System

Mail

Infrastructure Cost Estimator

From sensorsto intelligent systems.

Beyond personal projects.

Plexus

iNet Corporation

Tonirad Trans SRL

Tools, put to work.

Agents & retrieval

Vision & evaluation

Applications & delivery

Have somethingin mind?

Building AI.
From retrieval
to vision.

From sensors
to intelligent systems.

Have something
in mind?