Markus Clauss

AI & Data Science Lead · energy sector · since 2026

Dr. Markus Clauss (Abdullah Isa)

I lead AI and data science work in the energy sector, and I build software with AI coding agents: the written spec is the source of truth, the agent writes the code.

Dhahran, Kingdom of Saudi Arabia

See the projects GitHub
  • 21 yearsprofessional experience since 2005
  • 13 yearsdata science, ML and AI since 2013
  • Dr. rer. pol.economist, tax-benefit microsimulation and CGE modelling
  • 3 countriesGermany, Switzerland, now at home in Saudi Arabia

How I work

The documents are the source of truth. The agent writes the code.

Most of what I build today, I build with AI coding agents. Not by asking for snippets, but by writing the system down first and letting the agent implement it, one spec at a time.

Before a line of code exists, the repository holds the documents that define the system:

  • vision
  • architecture
  • ADRs
  • threat model
  • acceptance tests

The vision says why the thing exists. The architecture says how it is shaped. Each architecture decision record keeps a choice and its trade-offs. The threat model says what must never happen. The acceptance tests say what done means. The code is derived from these, and when the agent cannot implement a spec cleanly, the spec is what gets fixed.

I do this for the joy of it, and for real. The same discipline carries into my work with LLM systems on air-gapped GPU infrastructure: evaluation before rollout, governance written down, nothing that phones home.

  1. Write the spec

    Vision, architecture, ADRs, threat model and acceptance tests, in the repo next to the code.

  2. Hand it to the agent

    Claude Code implements one spec at a time. I read the diff the way I would read a colleague's.

  3. Run it for real

    Release candidates go on a VPS, a Mac and a phone. The acceptance tests decide, not the demo.

  4. Fold it back

    What the running system teaches goes into the documents first. The spec changes before the code does.

Why the names come from the Sira

My projects take their names from moments in the Sira, the life of the Prophet Muhammad ﷺ. Each name is chosen because that moment explains what the software is for. Thawr is the cave that gave shelter during the Hijra, so Thawr is a network that shelters. The names are meaning, not decoration.

I am at home in the Kingdom of Saudi Arabia now, and I am learning Arabic. Some of the smaller tools here exist because of that.

Projects

Built spec by spec, with agents

Thawr mark: a stylised cave arch in navy and teal

One binary. No cloud. Works offline.

  • Go
  • WireGuard
  • mesh VPN
  • ZTNA
  • SQLite
  • Apache-2.0
  • Claude Code

Thawr ثور

A self-hosted WireGuard mesh VPN and zero-trust network, written in Go. One binary runs as the control server, as the client on every device and as the admin CLI. State is a single SQLite directory. Policy is a YAML file kept in git, default deny. Phones join through the official WireGuard app. Nothing phones home, and it starts and runs with no internet access.

Named after the cave near Makkah where the Prophet ﷺ and Abu Bakr found shelter during the Hijra: hidden, protected, reachable only by those who belong. That is the design goal of the network.

Thawr is built spec by spec with Claude Code. The design documents, from vision to threat model and acceptance tests, live in the repository, and the implementation follows them one at a time. Release candidates run on a VPS, a Mac and a phone.

Source on GitHub Specs and design documents

AI agents & LLM systems

news-recap-app

AI News Recap: agents gather the day's topics, summarise them and turn the summary into a short video from generated images.

Second place, Encode Club / SwarmZero AI and Video Hackathon, December 2024.

  • multi-agent
  • Python
  • Next.js

Building with AI

arabictutor

Small tools for my own Arabic learning, built with agents as I go. A place to try ideas quickly and keep what helps.

Learning the language of my new home, one tool at a time.

  • Arabic
  • LLM
  • learning

AI research & evaluation

cache-implementation-challenge

Claude, Qwen and GLM-4.5 get the same specification and write the same Rust cache: LRU eviction, TTL, thread safety. The implementations are benchmarked side by side.

  • Rust
  • LLM evaluation
  • benchmarks

AI research & evaluation

apertus-transparency-guide

A transparency dashboard for Apertus, Switzerland's open 8B language model: attention patterns, weights and next-token probabilities, inspected live.

  • Apertus
  • Gradio
  • PyTorch

ML & data science

gradio-medical-image-analyzer

A Gradio custom component for veterinary medical image analysis with DICOM support, built so that AI agents can drive it.

  • computer vision
  • DICOM
  • Gradio

Background

Economist by training, engineer by practice

Career

  1. AI & Data Science Lead, Dhahran

    Leading AI and data science work in the energy sector: LLM systems and agents on air-gapped GPU infrastructure, evaluation and governance.

  2. Senior Data Scientist, University of Zurich (DIRU)

    Medical imaging AI and computer vision.

  3. Data Scientist / Data Engineer, Volkswagen / Autovision

    Advanced analytics, big data and cloud platforms.

  4. Data Scientist / Data Engineer, Sky Deutschland

    First years in data science and machine learning on production data.

  5. Software Developer, DATEV

    Professional software development.

  6. Researcher, ZEW Mannheim

    Labour markets and social security. Tax-benefit microsimulation, peer-reviewed publications.

Education

  1. Dr. rer. pol., University of Duisburg-Essen

    Tax-benefit microsimulation combined with computable general equilibrium modelling.

  2. Diplom-Volkswirt, University of Regensburg

    Econometrics.

Expertise

AI agents & LLM systems

Multi-agent systems, RAG, tool use, LLMOps, evaluation and governance, air-gapped GPU infrastructure.

ML & deep learning

PyTorch, TensorFlow, scikit-learn, XGBoost. Computer vision, NLP, forecasting.

Data & platform engineering

PySpark, Databricks, Delta Lake, Snowflake, dbt, Kubernetes.

Cloud & operations

AWS, Azure, GCP. Docker, Terraform, MLOps, CI/CD.

Languages

Python, R, SQL, Rust, Go, TypeScript.

Awards

  • Winner, Cudis-RADAR AI Hackathon (October). Shell AI team: a computer-vision ring size fitter built on HandMesh.
  • Second place, Encode Club / SwarmZero AI and Video Hackathon (December), with AI News Recap.
  • ZEW Qualification Award.

Where I come from

German national. I lived and worked in Germany and Switzerland and am now at home in the Kingdom of Saudi Arabia.

Publications

Peer-reviewed articles

  1. Franz, W., Gürtzgen, N., Schubert, S., Clauss, M. (2012). Assessing the Employment Effects of the German Welfare Reform: An Integrated CGE-Microsimulation Approach. Applied Economics, 44(19). doi:10.1080/00036846.2011.564149
  2. Clauss, M., Schnabel, R. (2008). Distributional and Behavioural Effects of the German Labour Market Reform. Journal for Labour Market Research (Zeitschrift für ArbeitsmarktForschung), 41(4). Full text (IAB)

Working papers and reports

  • Horstschräer, J., Clauss, M., Schnabel, R. (2010). An Unconditional Basic Income in the Family Context: Labor Supply and Distributional Effects. ZEW Discussion Paper 10-091.
  • Clauss, M., Schubert, S. (2009). The ZEW Combined Microsimulation-CGE Model: Innovative Tool for Applied Policy Analysis. ZEW Discussion Paper 09-062.
  • Clauss, M., Schnabel, R. (2008). Distributional and Behavioural Effects of the German Labour Market Reform. ZEW Discussion Paper 08-006.
  • Franz, W., Gürtzgen, N., Schubert, S., Clauss, M. (2007). Reformen im Niedriglohnsektor: eine integrierte CGE-Mikrosimulationsstudie der Arbeitsangebots- und Beschäftigungseffekte. ZEW Discussion Paper 07-085.
  • Arntz, M., Clauss, M., Kraus, M., Schnabel, R., Spermann, A., Wiemers, J. (2007). Arbeitsangebotseffekte und Verteilungswirkungen der Hartz-IV-Reform. IAB-Forschungsbericht 10/2007.
  • Contribution to ZEW Wachstums- und Konjunkturanalysen, 2006.

RePEc author page: link to be added

Contact

Get in touch

Interested in AI agents in regulated or air-gapped environments, spec-driven development, or Thawr? Write to me.