What Drives Me
I like learning how intelligent systems work: how data is prepared, how models learn, and how results improve through iteration.
// README.md
Marin Janushaj
- # Marin Janushaj
- role: AI Developer
- school: Thomas More University
- location: Geel, Belgium
- focus: machine learning, data, NLP, model training
- ## Main interests
- - training models and improving results
- - data handling and data engineering
- - intelligent systems and NLP
- ## Review path
- projects.case -> selected work
- stack.json -> technical range
- proof.md / contact.md -> next step
This portfolio is kept as a small workspace: start with the AI profile, inspect the project evidence, then open proof or contact without searching.
Education
Applied Computer Science at Thomas More University from 2023 to 2026, with strong school preparation in Tirana.
Languages
Albanian native, English IELTS 7.5, and B1 German, Turkish, and French for international environments.
// about.me.md
About Me
A deeper view of who I am, what I am focused on, and how I want to grow as an AI Developer.
AI Developer and Applied Computer Science student
I am based in Geel, Belgium, studying Applied Computer Science at Thomas More University. My main focus is artificial intelligence, machine learning, deep learning, data, natural language processing, and model training.
I like work that connects data, logic, and real people
AI interests me because it is not only about writing code. It is about preparing data well, testing assumptions, understanding why results change, and turning complex behavior into something useful and understandable.
Curious, practical, and improvement-driven
I like understanding how models learn, why results change, and how data, experimentation, and feedback can make intelligent systems better. I prefer improving through small iterations instead of trying to make the first version perfect.
Programming, data, BI, web, and AI technologies
My background includes Python, SQL, Java, JavaScript, PHP, Laravel, databases, business intelligence, data engineering, cloud AI tools, and application security basics.
Structured, open to feedback, and comfortable learning fast
I work best when the goal is clear, the feedback loop is honest, and there is space to test ideas. I like breaking large problems into smaller steps and improving the result with evidence.
Languages, sports, travel, and cultures
I value communication and international environments. I speak Albanian natively, English at C1 with IELTS 7.5, and German, Turkish, and French at B1 level. Sports, travel, and different cultures help me stay curious and adaptable.
Grow through real AI and data work
My goal is to gain more experience with machine learning and data projects, especially work where models, data quality, and product thinking come together. I want to help build intelligent systems that are useful, understandable, and well made.
// projects.case
Selected work
Six projects that support the AI Developer direction: agents, local NLP, machine learning systems, data, analytics, UX, and useful web interfaces.
ai_case
AI Learning Assistant
A multi-agent adaptive tutoring platform with 5 specialized agents, RAG over uploaded documents, student memory, feedback adaptation, and streaming chat.
- Role
- Multi-agent architecture, backend, RAG, and frontend integration
- Result
- A production-style AI tutor that adapts to how each student learns
Python
CrewAI
Gemini
FastAPI
React
Qdrant
Redis
Project Overview
A student chats with one assistant, while an orchestrator routes each question to a tutor, content curator, quiz generator, or exercise generator.
My Contributions
Built the FastAPI backend, multi-agent flow, RAG pipeline, Qdrant document search, Redis session memory, React interface, and Dockerized infrastructure.
railway_ai_case
TrackLine AI
A GDPR-compliant railway dispatching assistant, also framed as TrackGuard, that turns dispatcher-driver radio communication into classified, urgency-ranked incident reports, fully on-device.
- Role
- AI Lead for the speech-and-language module in a 7-person cross-disciplinary team
- Result
- Local Whisper transcription, diarization, semantic incident classification, entity extraction, urgency scoring, and report generation behind a FastAPI service
Whisper
pyannote.audio
FastAPI
ChromaDB
LLaMA 3.2 LoRA
Streamlit
Docker
Project Overview
TrackLine helps railway dispatchers process emergency radio traffic by transcribing audio, separating speakers, classifying incidents, extracting entities, and generating structured reports.
My Contributions
I owned the AI module: Whisper, two diarization approaches, semantic classification with ChromaDB, urgency scoring, FastAPI endpoints, and the Streamlit dashboard.
mlops_case
ML Prediction Systems
Two production-ready prediction systems: UK housing price prediction and UK electricity demand forecasting, built with LightGBM and deployed with cloud infrastructure.
- Role
- Housing pipeline, feature engineering, model training, deployment architecture, Docker, Oracle VM setup
- Result
- 22.5M housing transactions processed and electricity demand forecasting with 98.8% model accuracy
Python
LightGBM
Go
Flask
Docker
Oracle Cloud
Streamlit
Project Overview
A pair of deployable ML systems for regression and time-series forecasting, wrapped in APIs and Streamlit applications.
My Contributions
I focused on the housing model pipeline, feature engineering, training, deployment architecture, Docker orchestration, and Oracle Cloud setup.
data_case
Airbnb Dashboard
Business intelligence dashboard turning Airbnb listing data into pricing and booking insights.
- Role
- Data cleaning, analysis, and dashboard logic
- Result
- Interactive analytics view for host decision-making
Python
Data Analysis
Business Intelligence
Qlik
Project Overview
A data visualization project using raw Airbnb data to surface pricing and listing patterns.
My Contributions
Cleaned data, explored patterns, and shaped dashboard views around practical business questions.
supporting_web_case
Campus Bridge
Laravel and Livewire platform for incoming Thomas More students to find housing and programmes.
- Role
- Full-stack Laravel implementation
- Result
- A student support platform for moving to campus
Laravel
Livewire
MySQL
Tailwind
Project Overview
A student-support platform that centralizes housing, programmes, and practical information for incoming Thomas More students.
My Contributions
Worked on Laravel functionality, database structure, interface implementation, and project collaboration.
ux_case
SKIL2 Training Sessions
Research and high-fidelity Figma prototype for a training sessions app at Thomas More.
- Role
- User research and prototyping
- Result
- 40+ screen interactive prototype
Figma
UX Research
Prototyping
Testing
Project Overview
A product-design project focused on understanding users and translating research into clear app flows.
My Contributions
Worked through research, journeys, wireframes, prototype screens, and usability feedback.
// Realisation Paper Marin Janushaj.pdf
Realisation Paper Marin Janushaj
The main internship document for the Smart NV / SmartGames project. This is the primary file in the Internship folder.
main_pdf
Realisation Paper Marin Janushaj
A 52-page realisation paper documenting the SmartGames internship project: analysis, implementation, evaluation, results, and the final technical reflection.
- File
- Realisation Paper Marin Janushaj.pdf
- Main topic
- IQ Puzzler Pro digital companion with YOLO segmentation, board-state reconstruction, solver validation, hints, solving, and Remix export
- Open
- Use the button below to open the PDF in a new browser tab
What the paper proves
The project became more than a model demo. It connects phone capture, computer vision, deterministic puzzle logic, debug evidence, manual correction, and product-style UI flows.
Technical lesson
AI should not be trusted alone in a rules-based product. The model proposes masks, while board geometry and the exact-cover solver verify whether the state is actually playable.
// Project Plan Smart NV IQ Puzzler Pro.pdf
Project Plan Smart NV IQ Puzzler Pro
The planning document for the Smart NV internship project: scope, goals, technical approach, risks, phases, and success criteria.
planning_pdf
Project Plan Smart NV IQ Puzzler Pro
An updated project planning paper for the IQ Puzzler Pro computer vision puzzle tracking system, aligned with the final realised internship scope.
- File
- Project Plan Smart NV IQ Puzzler Pro.pdf
- Main topic
- Planning the IQ Puzzler Pro companion: front-board recognition, back-board recognition, synthetic data, solver logic, Remix, risks, phases, and evaluation
- Open
- Use the button below to open the PDF in a new browser tab
What the plan explains
It translates the internship into a structured planning document: problem, client context, project boundaries, technical approach, dependencies, risks, success criteria, and future roadmap.
Why it matters
The Realisation Paper shows what was built. This plan shows how the work was scoped and organised before and during implementation.
// Final reflection.pdf
Final Reflection
The closing reflection for the Smart NV internship: what was achieved, what still remains, and what changed in the way I work.
reflection_pdf
Final Reflection
A 5-page reflection on the Smart NV internship, covering the IQ Puzzler Pro computer-vision prototype, client value, remaining work, and personal growth.
- File
- Final reflection.pdf
- Main topic
- Substantive and personal reflection on the Smart NV computer-vision puzzle tracking system
- Open
- Use the button below to open the PDF in a new browser tab
What the reflection covers
It looks back on the delivered front-board workflow, back-board workflow, Remix prototype, YOLO segmentation, board-state reconstruction, validation, hints, solving, and STL export work.
Why it closes the folder
The plan explains how the work was scoped. The paper explains what was built. The reflection explains what was learned and what should happen next.
// internship/summary.md
Project summary
A concise explanation of the SmartGames internship project, separate from Selected work.
Smart NV / SmartGames internship
From February to May 2026, the assignment explored how IQ Puzzler Pro could be supported by a digital companion that recognises a real board and helps the player continue.
Photo to validated puzzle state
The flow combines phone capture, YOLO segmentation, OpenCV perspective correction, exact 55-cell mapping, and an exact-cover solver for validation, hints, and solving.
Front board, back board, and Remix
The delivered direction included front-board recognition, back-board recognition, debug screens, manual correction, and a Remix prototype for generated shapes, printable material, and STL export.
// internship/stack.json
Technical stack
{
"backend": ["Python", "FastAPI"],
"computerVision": ["OpenCV", "NumPy", "YOLO v26 nano segmentation"],
"machineLearning": ["Ultralytics YOLO", "PyTorch", "ONNX"],
"frontend": ["React", "Vite", "TypeScript"],
"dataGeneration": ["Blender Python", "Synthetic data", "Automatic labels"],
"logic": ["55-cell board state", "Exact-cover solver", "Hint and solve endpoints"],
"exports": ["PDF booklet", "SVG", "STL"]
}
// learning.case
AI Learning Assistant
A multi-agent adaptive tutoring platform with RAG, session memory, feedback adaptation, and production-style infrastructure.
multi_agent_system
AI Learning Assistant — Multi-Agent Adaptive Tutoring Platform
A student chats with it like ChatGPT, but behind the scenes an Orchestrator agent analyzes the question and routes it to the right specialist.
- Agents
- Orchestrator, Tutor, Content Curator, Quiz Generator, Exercise Generator
- Memory
- Qdrant for RAG over uploaded materials, Redis for conversation context
- Adaptation
- Thumbs-up/down feedback and quiz performance adjust difficulty and teaching style
Python
CrewAI
Google Gemini
FastAPI
React
Qdrant
Redis
Docker
What makes it more than a chatbot wrapper
It uses intent routing, document-grounded answers, adaptive teaching, real-time web curation, session persistence, streaming responses, and Dockerized infrastructure.
Portfolio one-liner
A multi-agent AI tutor that learns how you learn — RAG, adaptive difficulty, and 5 specialized agents behind one chat box.
// trackline-ai.case
TrackLine AI
A privacy-first railway incident detection system built in a 7-person cross-disciplinary team, with my ownership focused on the AI speech-and-language module.
local_nlp_pipeline
TrackLine AI — Real-Time Railway Incident Detection
A GDPR-compliant dispatcher assistant that listens to railway radio traffic and turns it into structured, urgency-ranked incident data in seconds.
- Team
- 7-person cross-disciplinary team across AI, app development, and cybersecurity
- My module
- Whisper transcription, diarization comparison, semantic classification, entity extraction, urgency scoring, FastAPI service, and Streamlit dashboard
- Constraint
- 100% local inference for GDPR, with no cloud processing and audio deleted after analysis
Whisper
pyannote.audio
sentence-transformers
ChromaDB
LLaMA 3.2 LoRA
FastAPI
Streamlit
Docker
What makes it strong
I built the AI service that the .NET backend and React frontend integrated against, which meant designing stable API contracts and making the AI output usable for the rest of the product.
Engineering trade-off
The module includes two diarization paths: a fast railway-keyword heuristic around 1 second and a more accurate pyannote.audio voice-based path around 5 seconds, exposed through a comparison endpoint.
// ml-prediction-systems.case
Two Production-Ready Prediction Systems
An end-to-end machine learning project covering UK housing price prediction and UK electricity demand forecasting.
prediction_systems
Housing Price Prediction + Electricity Demand Forecasting
Built with my teammate Yunus Eren ERTAS: two deployable ML systems using LightGBM, Streamlit interfaces, Dockerized services, and cloud-hosted infrastructure.
- Housing
- 22.5M transactions from 1995-2017, R2 = 0.684, average error GBP 61,409, trained in 27 seconds
- Electricity
- 25 years of NESO data from 2001-2025, R2 = 0.988, RMSE 595 MW
- Architecture
- Go API Gateway, Python Flask ML Service, Docker orchestration, Oracle Cloud VM, Streamlit apps
Python
LightGBM
Go API Gateway
Flask
Streamlit
Docker
Oracle Cloud
My focus
Housing dataset pipeline, feature engineering, model training, deployment architecture, Docker orchestration, and Oracle VM setup.
Why it matters
The systems show applied ML beyond notebooks: full-data training, feature engineering, APIs, deployment, and usable prediction interfaces.
// airbnb-dashboard.case
Airbnb Dashboard
A business intelligence project turning listing data into practical pricing and booking insight.
dashboard_system
Pricing, booking, and listing signals
The dashboard organizes Airbnb data into a clear analytical view so patterns become easier to inspect and explain.
- Role
- Data cleaning, exploration, dashboard logic, and insight framing
- Focus
- Turn raw listings into decisions around price, location, and host performance
Python
Data Analysis
Business Intelligence
Qlik
// campus-bridge.case
Campus Bridge
A Laravel and Livewire platform helping incoming Thomas More students find housing, programmes, and support.
web_platform
Student support, structured like a product
The project centralizes practical information for students arriving at campus, with clear flows for housing and programme discovery.
- Role
- Full-stack Laravel implementation, database structure, interface work, and collaboration
- Result
- A useful student-support platform that turns scattered information into a guided experience
Laravel
Livewire
MySQL
Tailwind
// skil2-training.case
SKIL2 Training Sessions
A UX research and prototyping project for a training sessions app at Thomas More.
prototype_system
Research translated into a 40+ screen prototype
The work moved from user research and journeys into wireframes, high-fidelity screens, and usability feedback.
- Role
- User research, journey mapping, wireframing, high-fidelity prototyping, and iteration
- Result
- A clear app flow that shows product thinking, usability, and interface discipline
Figma
UX Research
Prototyping
Testing
// stack.json
Stack
AI, data, programming, cloud tools, languages, and soft skills that support AI/data-related opportunities.
{
"aiData": ["Machine Learning", "Deep Learning", "NLP", "Business Intelligence", "Data Engineering", "Model Training", "Qdrant", "SQL"],
"programming": ["Python", "Java", "JavaScript", "PHP", ".NET", "HTML", "CSS", "Sass", "React", "Laravel", "Livewire"],
"cloudTools": ["AWS", "Cloud AI platforms", "Scrum", "Figma", "Application security basics"],
"languages": ["Albanian native", "English C1 / IELTS 7.5", "German B1", "Turkish B1", "French B1"]
}
Machine learning, deep learning, NLP, data engineering, model training
The core direction is learning how models work, improving results, and applying AI to practical projects.
Languages, communication, cultures, fast learning
International communication and continuous learning are part of the profile, not side notes.
// proof.md
Proof
Education, experience, awards, and certifications kept close to the work for quick review.
Curriculum Vitae
PDF
Updated one-page CV with education, IT skills, activities, awards, and contact details.
A Level Computer Science Course
PDF
Certified course connected to earlier Computer Science preparation.
ITirana Shadowing Proof
Business and Informatics Technology company experience, 2022 - 2023.
AI Internship at SMART NV
February 2026 - May 2026, focused on AI-related professional experience.
// contact.md
Contact
The fastest way to reach me for AI, data, internship, or collaboration opportunities is email or LinkedIn.