say hi
~ /home/lukas · available 2026

Lukas
Mührke

engineer·researcher·maker

I build and study machine learning systems at the boundary of theory and deployment — where inference, sequence models, and hardware interact under real constraints.

↓ scroll to explore→ dots to jump
— 02 / about

A few things
about me.

I came to computer science via boatbuilding — three years at a racing yacht yard in Kiel before moving into a CS degree at TU Hamburg. That background still shapes how I think: materials have constraints, structures have failure modes, and the most interesting problems sit at the boundary between what is theoretically possible and what actually holds together.

My work focuses on machine learning systems in real environments — models that operate under latency budgets, on constrained hardware, and with imperfect or shifting signals. I’m interested in how architectural choices propagate into behavior once a model leaves the lab, and how this behavior can be made stable, predictable, and understandable.

A recurring theme in my work is sequence modeling as an alternative to attention in domains where its inductive biases better match the structure of the problem. More broadly, I’m interested in how representations, constraints, and system design interact — and how small changes at one level shape outcomes at another.

I’m currently working on inference and system-level questions in applied settings, including model optimization and multi-agent architectures, with experience spanning both industrial research and applied deployments.

Alongside that, I care about how these systems present themselves — how interfaces, feedback, and interaction design influence whether a system feels reliable or opaque.

Outside of work, I run, read more than I should, and keep a paper notebook that doesn’t sync anywhere.

Based in
Hamburg, DE
Studying
M.Sc. Computer Science, TUHH
Working on
SSM research · inference infra · hardware prototypes
Available
select collaborations · 2026
— 03 / skills

What I reach for first.

— build
  • Python
  • TypeScript
  • C/C++
  • Kotlin
  • Bash
— craft
  • Fusion 360
  • Bambu P1S / FDM
  • CFK/GFK composites
  • KiCad
— think
  • ML systems
  • LLM inference & optimization
  • Sequence modeling
  • Multi-agent systems
  • Research
— 04 / projects

Some things I've built.

— drag to scroll →
nrc-mcp
Unofficial Nike Run Club MCP server — exposes run history and stats to LLM agents via Bearer auth.
// python · mcp
postattention
Research comparing SSM-based models (Mamba, S4) against attention baselines for CNC tool-wear prediction.
// pytorch · wandb
inpainting-fft
BSc thesis: image inpainting with a frequency-domain loss using Fast Fourier Transforms.
// python · pytorch
clip-labeler
Hierarchical image classification pipeline using CLIP and DINO for a WordNet-scale generation dataset.
// python · sqlite
lab-psu
Modular 30V/10A DC bench supply with ESP32-S3 control and a MOSFET switching matrix.
// esp32 · kicad · c
— 05 / cv

A short timeline.

  1. 2025—
    AI Research Working Student
    Cluster management, inference optimization, multi-agent system development.
    NXP Semiconductors
  2. 2024—
    M.Sc. Computer Science
    Focus: ML systems, embedded systems, high-performance computing, system security.
    TU Hamburg
  3. 2024–25
    Student Research Assistant
    Transcription model fine-tuning, LLM-based radio message processing, TTS for synthetic audio.
    Fraunhofer CML
  4. 2018–24
    B.Sc. Computer Science
    Thesis: image inpainting using deep learning with a frequency-domain loss.
    TU Hamburg
  5. 2020–22
    Working Student QA
    ePages GmbH
  6. 2015–18
    Boatbuilding Apprentice
    CFK/GFK composites, precision fabrication, racing yacht construction.
    Knierim Yachtbau
— 06 / contact

Say hi.

If any of this resonates — or you just want to trade notes — I'd like to hear from you.

© 2026 · Hamburg, DE · made with care