J Jakub Jaromin
CV
Jakub Jaromin

Jakub Jaromin

I work at the intersection of artificial intelligence and mathematical physics. At the Faculty of Applied Mathematics (B.Sc. Computer Science, specialization in AI & Data Analysis) I contribute to research articles, maintain laboratory computing infrastructure, and build tools that make research easier.

My thesis in preparation trains a physics-informed neural network to satisfy Einstein’s vacuum equations for a rotating black hole, then uses the learned metric as a geometry engine for relativistic ray tracing. Beyond general relativity, I work on geometric data augmentation, audio signal analysis, satellite imagery, and medical imaging.

AI · General Relativity · Signals · Quantum · Mathematics

Projects

Problem, approach, and result — with code where it is public.

Live Tonation Detection interface

Signal processing · audio

Live Tonation Detection

RAG Research Assistant home page

LLM tooling · research

RAG Research Assistant

Satellite change detection sample output

Computer vision · remote sensing

Satellite Change Detection

FOBO medical imaging dashboard

Medical imaging · deep learning

FOBO — Lung & Heart Analysis

n8n agent workflow graph

LLM automation · workflows

AI Task Assignment Agent

Physics-informed deep learning · general relativity

PINN-Kerr — Black Hole Solver & Renderer

Problem
Can a neural network learn an exact solution of Einstein's vacuum field equations — the Kerr metric of a rotating black hole — well enough to be used as a spacetime geometry engine?
Approach
A physics-informed network (JAX) learns the Kerr–Schild scalar H(x, y, z, a) in Cartesian Kerr–Schild coordinates, conditioned on spin a ∈ [0, 0.99]. Four physics-motivated loss terms (Ricci residual, analytic anchor, Sobolev smoothness, anti-Minkowski) are combined, with Christoffel, Riemann, and Ricci tensors computed by automatic differentiation. Hyperparameters tuned with Optuna; experiments tracked in MLflow.
Result
The trained network renders photorealistic black hole images by integrating null geodesics through the learned metric, with a Novikov–Thorne accretion disk model — renders up to 8000×4000. Basis of my engineering thesis (in preparation).
  • JAX
  • PINN
  • General Relativity
  • Ray Tracing
  • Optuna
  • MLflow
Kerr black hole with accretion disk, rendered by ray tracing through the learned metric

Signal processing · audio

Live Tonation Detection

Problem
Identify the musical key (major/minor) of live audio in real time.
Approach
Fourier-transform analysis of the live audio stream with a lightweight desktop interface (Python, Tkinter) for monitoring predictions.
Result
Working real-time key detector with a live interface; code public on GitHub.
  • Python
  • FFT
  • Audio
  • Tkinter
Live Tonation Detection interface Code fragment of the tonation detector
Repository

LLM tooling · research

RAG Research Assistant

Problem
Reading research PDFs is slow; finding answers across papers is slower.
Approach
Retrieval-Augmented Generation pipeline over PDF documents fetched from arXiv: chunking, embedding, retrieval, and a chat interface with a configuration page.
Result
Chat-with-papers assistant for arXiv documents; code public on GitHub.
  • LLM
  • RAG
  • Python
  • arXiv
RAG Research Assistant home page Configuration page of the assistant
Repository

Computer vision · remote sensing

Satellite Change Detection

Problem
Detect changes between satellite images of the same area over time, for urban and environmental monitoring.
Approach
Siamese deep learning architectures, including FC-Siam-Conc, trained on high-resolution satellite imagery (PyTorch).
Result
Trained change-detection models with qualitative results on high-resolution scenes.
  • PyTorch
  • Deep Learning
  • Satellite
  • Change Detection
Satellite change detection sample output Training samples from the JL1 datasetFC-Siam-Conc architecture diagram

Medical imaging · deep learning

FOBO — Lung & Heart Analysis

Problem
Support X-ray diagnostics by localizing lung and heart regions of interest.
Approach
Deep learning on medical X-ray images generating lung and heart heatmaps for diagnostic assistance.
Result
Heatmap pipeline highlighting cardiac and pulmonary regions on chest X-rays.
  • Medical AI
  • Deep Learning
  • Heatmaps
  • Python
FOBO medical imaging dashboard Lung and heart heatmap on a chest X-ray

LLM automation · workflows

AI Task Assignment Agent

Problem
Manual task planning and assignment across team tools wastes time.
Approach
An n8n workflow agent driven by an LLM (Gemini, LangChain) that integrates Google Sheets and Docs to assign tasks automatically.
Result
Working automation pipeline assigning tasks from spreadsheet data; code public on GitHub.
  • n8n
  • LLM
  • LangChain
  • Automation
n8n agent workflow graph Task board integration view
Repository

Research

Research

Peer-reviewed articles and thesis work.

RAAM point cloud visualization on a cancer dataset

Under review

Radial-Angle Augmentation Method (RAAM)

A geometric data augmentation technique for point sets. The point cloud is centred on its centroid, rotated so the farthest sample defines the first angular segment, and divided into circular segments of angular width α with equal-area subsections. Synthetic points are distributed proportionally to the existing point density per subsection, preserving the statistical structure of the original cloud.

Code fragment of the RAAM implementation
  • Point Cloud
  • Geometric Augmentation
  • Polar Coordinates
  • Density Preservation
Card recognition model reaching 98.5% accuracy

Accepted — in press

Card Recognition with Neural Networks

Playing card recognition with convolutional neural networks: a custom-trained model reaching 98.5% accuracy, with analysis of training metrics and confusion matrices for color classification.

Model accuracy across training epochsConfusion matrix for card color classification
  • CNN
  • Computer Vision
  • Image Classification
Kerr black hole render from the physics-informed neural network pipeline

Thesis

Physics-Informed Neural Networks for the Kerr Metric

Training a neural network to satisfy Einstein's vacuum equations for a rotating black hole in Cartesian Kerr–Schild coordinates, and using the learned metric as a geometry engine for null-geodesic ray tracing of photorealistic black hole images.

  • PINN
  • General Relativity
  • JAX
  • Ray Tracing

Experience

Experience & education

Education

Silesian University of Technology

B.Sc. Computer Science — Faculty of Applied Mathematics

Specialization: AI & Data Analysis

Grant

“Najlepsi z najlepszych! 4.0”

Polish Ministry of Education and Science, 2025

Grant programme supporting outstanding students with exceptional scientific achievements and active participation in research projects.

gov.pl

Jun 2026 — present

Intern

KP Labs

Internship at KP Labs, a space technology company.

Oct 2025 — present

Laboratory Technician

Silesian University of Technology

Maintaining and developing computing infrastructure for research and teaching; applied studies in machine learning, data visualization, and signal analysis.

Dec 2024 — present

Research Collaboration

Faculty of Applied Mathematics, SUT

Contributing to scientific articles within a research team.

Oct 2024 — present

Member

Data Science Scientific Circle, SUT

Workshops, hackathons, and group projects in machine learning, data visualization, and statistical analysis.

Activity

Hackathons, workshops, and science outreach.

HackYeah 2025 award ceremony on stage Rest & Blink statistics window of the winning app

2025 · 1st place — Biohacking

HackYeah 2025

Won the Biohacking category at Europe’s largest stationary hackathon — the most competitive track, with nearly 60 international teams. Built “Rest & Blink”, an AI wellness app using computer vision and heart-rate monitoring.

SGH × Mastercard hackathon team photo Heat map of areas where a parcel locker is needed

2025 · Finalist — top 8 of 140

SGH × Mastercard Hackathon

Reached the finals (8 of 140 teams) with a geospatial scoring model and prototype web app identifying optimal parcel locker locations.

Neural network workshop session

2024 · Workshop lead

Neural Network Workshop

Led and co-ran workshops on neural network fundamentals — from the mathematics of perceptrons to convolutional architectures.

Data Science Circle booth at the Silesian Science Festival

2024 · Science outreach

Silesian Science Festival

Represented the Data Science Scientific Circle at a public booth, talking data science and AI with festival visitors.

Meeting with astronaut Sławosz Uznański

2024 · University visit

Meeting Sławosz Uznański

Attended a meeting with the ESA project astronaut during his university visit — space exploration and technology first-hand.

Contact

Open to research collaboration, PhD opportunities, and technical work. Email or GitHub is the fastest path.

Email [email protected] GitHub github.com/Jaromek LinkedIn linkedin.com/in/jaromek ORCID 0009-0006-0664-1580 Download CV (PDF)