// silkscreen
I build systems from first principles — from copper traces and custom silicon up to transformers and RAG pipelines.
Software engineer specialising in machine learning and automation, with a strong foundation in RAG systems and embedded / hardware development. Currently studying Software Engineering at TU Wien while building machine-learning and automation tooling at ORF, Austria's national public broadcaster. Comfortable across the entire stack — neural networks and transformer models on one end, microcontrollers, RF engineering and a taped-out chip on the other. If I use something, I usually end up rebuilding it from scratch to understand it: CPUs, emulators, GPTs, mesh protocols.
| Stack depth | silicon → cloud |
| Lowest level shipped | SKY130 ASIC tape-out |
| Highest level shipped | RAG agents in production |
| Routing decision | ~15 µs (custom mesh protocol) |
| Operating location | Vienna, AT |
Machine-learning solutions for internal processes at Austria's national broadcaster: a multi-module RAG agent over company knowledge (SharePoint, wikis, enterprise architecture), transformer-based e-mail classification & routing (KIKO), and machine-learning-assisted video editing wired into Adobe Premiere (FlowCut). Deployed on AWS EKS with GitOps (ArgoCD, Crossplane, Helm).
Analysis and resolution of network issues in enterprise environments.
Features for automated timetable generation in school software; internal company website. C++ and Angular in a modern development team.
First contact with automation technology, machinery and programming.
Selected modules. Each one exists because I wanted to know how the thing underneath works.
Dual-band mesh network on ESP32-S3: LoRa 433/868 MHz as control plane, Wi-Fi HaLow (802.11ah) as data plane. Custom geographic routing protocol with ~15 µs decisions, GPS positioning, OTA updates, link-budget math and own antenna designs. Python simulation scaling to 50,000 nodes; Node.js/MongoDB backend with a Vue dashboard.
Humanoid with Adeept 5-DOF arms and tendon-driven InMoov hands, teleoperated live through a Meta Quest 3. Stereo global-shutter cameras on a Raspberry Pi 4 stream into the headset; a MoE-inspired action-tokenization architecture enables behavioral cloning from the recorded teleop data.
8-bit RISC CPU designed from scratch in Logisim/Digital with microcode ROM and Von Neumann architecture — then taken all the way to real silicon via TinyTapeout on the SKY130 process using OpenLane and Verilog.
RAG agent with scatter-gather retrieval across 17 services and modular knowledge sources (SharePoint, BookStack, wiki, enterprise architecture), backed by Qdrant with hybrid search. Fills out internal forms and automates service processes, not just Q&A.
Analyzes raw footage and assists editors via prompts or auto-generates promos and trailers. Hooked into Adobe Premiere Pro through a plugin driven by a Python server that writes cuts directly into the timeline.
QPSK and OFDM modulators written in Verilog on a Spartan-7, plus work with AD9361-class SDR transceivers, PCB Yagi/LPDA antenna design at 868 MHz / 2.4 GHz, VNA measurements and a custom signal generator.
NES emulator with Raylib, a CHIP-8 emulator, a mini GPT and a neural network — all written from scratch in C. The recurring theme: reimplement it until the magic is gone.
View on GitHub →Full pipeline in Python: terrain and heightmaps, tensor-field street networks, BSP building interiors with furniture and door connectivity. Rendered in a Three.js first-person viewer and validated with headless Node.js gameplay tests.
Led a four-person diploma team building armbands that read arm cross-section impedance plus IMU data and translate signs to speech in real time through a neural network — PyTorch and MATLAB models, custom GUI for recording, training and inference, output on a smartphone.
Self-built RC car driven in real time from a Meta Quest 3. ESP32 with ESP-IDF handles control and sensors; onboard camera streams low-latency video into the headset via WebXR with bidirectional comms.
Noninvasive BCI built around an ADS1299 analog front-end and an ESP32-S3, with BrainFlow for acquisition and processing. Ongoing exploration of the physical limits of noninvasive neural sensing.
4th place in a TU Wien cybersecurity CTF; solved challenges in blind SQLi, CSP nonce prediction, DNS spoofing and AES-ECB/CTR attacks. Built two agents in 24h at START Hack Vienna: an SSH service-desk autopilot and a voice cancellation-refill agent.
A native 3D renderer injected into Factorio — a 2D top-down game — turning it into a real-time 3D world. Reverse-engineered the game binary: a Rust DLL resolves and hooks engine functions by PDB symbol name, with RVA fallbacks that survive game updates, reads entity and world state directly from process memory, and draws the scene through custom HLSL/Direct3D shaders — orbit and first-person cameras, day/night lighting and live tuning, while menus, GUIs and mouse picking keep working.
Machine learning, algorithms & data structures, visual computing. 4th place in a cybersecurity CTF.
Circuit design, digital technology, natural sciences.
Graduated with merit. Diploma project: Sign-to-Speech (team lead, µC programming, ML).
Want to build something that goes deeper than the framework layer?