Portrait of Jakub Majewski. Software engineer
Munich

Jakub Majewski

I'm a software engineer in Munich with a background in Management & Technology. I build software end to end: XR apps, real-time 3D from engineering CAD, machine learning and knowledge tools for teams that use AI agents. At TUM I research Gaussian Splatting for real-time 3D on headsets.

GitHub LinkedIn contact@majewski.studio

Let's stay in touch

Selected work

Seven projects from research and industry, each with a full case study on GitHub

A gearbox, a jet engine, a Formula 1 car and a carbine rifle: four of the five 3D models in the study, rendered in Blender Cycles. — Four of the five study objects in Blender Cycles

Eight Gaussian splat models of an industrial plant on the headset, in passthrough. — Eight trained models on the headset

Research · TUM · 2026

Gaussian Splatting on Apple Vision Pro

Test bench for Gaussian Splatting on a standalone headset. It measures how close a trained model stays to the original 3D object and what it costs to render.

I wrote the image scoring in NumPy and a visionOS measurement app around a Metal splat renderer. The analysis plan was fixed before the results came in and every number can be rebuilt from the raw records.

At SplatFactory3D I work on large Gaussian Splatting scenes for standalone headsets. One scene I captured and trained myself is public on SuperSplat.

470: Gaussian splat models of five 3D objects, trained on cloud GPUs
14.1 million: training steps in total
203,520: per-view scores against reference renders
195: half-minute measurement runs on the headset with frame timing and thermal state

3D Gaussian Splatting visionOS Metal Python Experimental design

Read the case study

Renders of CC BY 4.0 models. Credits and changes are in the case study

Two hands holding the ends of a ruler across an aligned 3D dental scan. The ruler shows the distance in millimetres. — Measuring a dental scan by hand

The jaw bone and teeth from a CBCT scan, shown inside the aligned 3D model. — Jaw and teeth from the CBCT

Industry · Client project

receptionOS Visualizer

XR app that shows a patient's face scan, jaw scans and CBCT aligned in 3D, with hand-placed measurements in millimetres.

The client had the scans and the headset but no workflow to use them together. I built it as a hybrid: SwiftUI for the native app, Unity with URP and Metal for the 3D rendering. The ruler snaps to the scan surface and can measure between two aligned models, for example a jaw and its CBCT. I took it from an early proof of concept to the client release. Version 2.1.5 was accepted on the headset.

visionOS SwiftUI Unity C# Metal

Read the case study

Top-view schematic of stereo mirrors on visionOS: each eye gets its own reflected camera, with an off-axis frustum through the mirror outline. — One reflected camera per eye

Diagram of head-up display clipping in mixed reality, tested per pixel against the visible area of the windshield. — HUD clipping checked from 90 head positions

Industry · In-house application

Automotive XR

App for full-scale vehicle design reviews on a headset, with hand-tracked interaction and switching between variants.

I took it over from an earlier research prototype. I work on rendering, interaction and release engineering. Mirrors render a separate reflection for each eye because RealityKit has no planar reflections. Glass and HUD render in the correct order and the HUD is clipped per pixel against the windshield from both front seats. The steering wheel, doors and props are operated with tracked hands.

Unity PolySpatial RealityKit C#

Read the case study

One quilted leather material shown under four lighting presets: city midday, courtyard, night street and sunset. — One material under four lighting presets

An example part in six stages, from the original CAD through tessellation, reduction, repair and UVs to real-time materials. — From engineering CAD to real-time materials

Industry · From engineering CAD to real time

Automotive content pipeline

Geometry processing and tooling that turn heavy engineering CAD into real-time 3D for XR.

I integrated a CAD tessellation toolkit after an earlier attempt had left the editor crashing. Seat quilting is cut along the design lines directly in the CAD mesh and the stitches are real geometry. 138,254 collider triangles were replaced by 30 fitted collision shapes.

Python Blender Unity Asset Transformer (Pixyz) Unity

Read the case study

A graph of connected engineering knowledge: dense clusters of sources linked to decisions and checks. — Knowledge graph from the links between pages

Search time as the knowledge base grows, version 1.0 against 2.0

Tooling · Local engineering knowledge system · Version 2.0

LEKS

Local knowledge system for a long-running engineering project. Markdown pages and their source files in one folder, read by people in Obsidian and queried by AI coding agents over MCP.

In version 2.0 I rebuilt the layer under the search: a persistent BM25 index, hash-addressed copies of all 2,850 source files and claims linked to the lines they rely on. Rankings stay identical to 1.0 on all 941 test questions, page search is 176 times faster and peak memory of the full integrity check dropped from 1,602 MB to 82 MB.

Python Information retrieval MCP Obsidian

Read the case study

Predicted against true heating and cooling loads on held-out designs; the points follow the diagonal. — Held-out predictions for heating and cooling

Reliability curve comparing the stated interval coverage with how often the intervals actually contained the truth. — Calibration of the prediction intervals

Course project · Machine learning · TUM

Probabilistic building energy prediction

Heating and cooling load prediction for building designs with an uncertainty estimate for each prediction.

I compared ridge regression, a 300-tree random forest and exact Gaussian-process regression on the same held-out splits. The random forest was more accurate on heating load and the Gaussian process on cooling load. The Gaussian process also gives an uncertainty estimate, so the report keeps both models. Prediction intervals are checked for calibration and against split-conformal prediction. 11 automated tests run in CI.

Python scikit-learn Uncertainty quantification

Read the case study

Avatar tutor condition

The text tutor condition: an AI tutor chat that answers questions about the lesson. — Text tutor condition

Research · Human-AI interaction · TUM

AI Study Buddy

Platform for a user study that compares an AI text tutor with a speaking avatar tutor.

I was technical and product lead in a five-person student team and built the platform end to end. It covers consent, a baseline test, a learning session with one of the two tutors and a post-test, plus logging, flagging of invalid sessions and data export for analysis. On one key question correct answers went from 23% before to 95% after the session.

Human-AI interaction User study

Read the case study

About

I study Management & Technology at TUM with Computer Engineering as my technology specialization. Before that I did an M.Sc. in Management at Gdańsk University of Technology and got the faculty's award for the best master's thesis of 2024. I'm also Head of Strategic Development at TUM-XR, the university's student XR initiative, and a Senior Associate at XR Association Munich e.V.

Experience

Since 10/2025Software developer and technical lead for an automotive XR application
Since 09/2025XR developer for client projects
2023 to 2025Digital product developer for XR in industrial engineering

Publication

2023New challenges, new solutions: How does an IT company navigate in the VUCA era?With Krzysztof Leja · e-mentor 3(100) · in Polish

Tools

LanguagesPython, C#, Swift
XR and 3DUnity, RealityKit, PolySpatial, SwiftUI, Metal, Blender, Unity Asset Transformer (Pixyz)
Machine learningPyTorch, NumPy, scikit-learn
CloudGoogle Cloud

Contact

Let's stay in touch

contact@majewski.studio LinkedIn GitHub