Portrait of Paul Volk
Informatics · TU Munich

Agents, ML and the systems beneath them.

I build LLM agents, machine-learning systems and the infrastructure they run on.

GitHub paul.volk@tum.de ›

Projects

Self-Hosted Agent Platform

A central MCP gateway (FastMCP) that gives my LLM agents a shared toolbox – web search, scraping, documentation lookup, Google Workspace and Home Assistant – plus a shared long-term memory. Four different agents connect to the same tools and memory over SSE and streamable HTTP. BM25 tool search keeps their context small even with 100+ tools.

PythonFastMCPMCPDockerTraefikLLM agents

AI Photo Booth

A photo booth for our dorm's theme parties. A Raspberry Pi takes the photo, a FastAPI service detects faces and separates people from the background, and Stable Diffusion repaints only the background and clothes to match the theme – faces stay untouched. About 1–2 seconds per image on an RTX 2070 Super, printed on a label printer.

FastAPIComfyUIStable DiffusionInsightFaceRaspberry Pi

Jarvis – Voice Assistant

A custom voice assistant on a Raspberry Pi with Gemini as the LLM, local retrieval-augmented memory, Home Assistant integration and wake-word detection. Speech recognition and German text-to-speech run on a self-hosted GPU server. Open source.

PythonGeminiRAGfaster-whisper
View on GitHub ›
In progress

Speech to Protocol

An interdisciplinary project with TUM's Chair of Bioseparation Engineering for hands-free lab documentation: researchers dictate what they do, the system transcribes it, extracts structured data and files it in the lab's electronic lab notebook (eLabFTW). My part: grounded LLM-based protocol generation and the eLabFTW integration.

LLMsSpeech-to-texteLabFTW
In progress

Residence Drinks System

A drinks-tally system for the residence bar: residents pay with chips at scanner terminals in each house. Adapted from the system used at Studentenstadt Biederstein.

EmbeddedSelf-hosted

Personal Agent Automations

A morning briefing from calendar, tasks and mail, and a training coach that suggests the day's session and posts to Telegram – triggered when I actually wake up rather than at a fixed time. Plus a small web app for planning my weekly training.

OpenClawNode.jsMCPTelegram

Homelab

Proxmox VE with TrueNAS (ZFS) and a containerized Docker stack behind Traefik, with VPN access via Tailscale and WireGuard. Hosts the agent platform, the speech-to-text server and most of the projects above, plus Home Assistant, AdGuard and Ollama.

ProxmoxZFSDockerTraefikWireGuard

Small Hardware Builds

Arduino washing-machine monitors that use an accelerometer to show residents whether a machine is free, and an ESPHome IR blaster that controls an old Yamaha AV receiver from Home Assistant.

ArduinoESPHomeIoT

Experience

08/2026 – now

Network & System Administrator

Studentenwohnheim Maßmannplatz, Munich

  • UniFi LAN and Wi-Fi for five houses, with per-house VLANs to cut multicast traffic
  • Proxmox server running Nextcloud, Vaultwarden, Home Assistant and Jellyfin
  • Traefik reverse proxy with automated TLS certificates
10/2025 – 02/2026

NLP Research Project

TUM · Robustness of BERTScore, BLEURT and LLM-as-a-judge against semantic shifts and hallucinations

11/2024 – 12/2024

Student Research Assistant

Helmholtz Munich · Follow-up experiments on the thesis pipeline

06/2024 – 10/2024

Bachelor's Thesis

Helmholtz Munich · Weakly supervised instance segmentation of out-of-distribution red blood cell data

04/2023 – 09/2023

Medical Image Segmentation Project

CAMP, TUM · Nerve segmentation with U-Net, supervised by Prof. Dr. Nassir Navab

2022 – 2024

Volunteer Network Administrator

NetzwerkAG Biederstein · Student residence with ~170 residents

Education

04/2025 – now

M.Sc. Informatics

Technical University of Munich

09/2022 – 03/2023

Exchange semester

UC San Diego, USA

10/2019 – 03/2025

B.Sc. Informatics

Technical University of Munich

About

I'm an Informatics Master's student at TU Munich working on LLM agents and machine learning. My research so far covers weakly supervised instance segmentation and the robustness of text-generation metrics against hallucinations.

Outside of university I like building things people actually use: agents that share tools and memory, a voice assistant, an AI photo booth for dorm parties, and the network and servers of the student residence I live in.

Languages
Python, C++, Java, SQL, C, Haskell
ML / AI
PyTorch, Hugging Face, LLM APIs, MCP, RAG, LLM evaluation, image segmentation, Stable Diffusion
Systems
Linux, Docker, Proxmox, ZFS, Traefik, Git, UniFi, VLANs, WireGuard
Spoken languages
German (native), English (C1), French (B2)
Engagement
President of StudentInnen am Biederstein e.V. (2023–24) · Committee member, Studierendenwerk München (2023–24) · Network working group, Maßmannplatz
Interests
Climbing and bouldering, running, cycling, swimming, homelab and self-hosting