Blog posts
-
The open-source AI layer nobody talks about
While everyone’s watching the LLM race, another huge leap just happened — and hardly anyone is talking about it.
-
Document layout
Only once you know which region of a page is running text, which is a table, which is a diagram, and which is a caption — only then can you decompose the page into clean Markdown: linear text in reading order, plus a description for every figure.
-
Video search
I’ve always wanted to understand how you make videos properly searchable. So I built a small project with open models running locally on my laptop.
-
Knowing When You're Done Talking: What Building a Voice Pipeline Taught Me
Building a small dictation tool turned into a tour of every place a speech-to-speech pipeline has to guess when you're done talking -- and where that guessing goes wrong. -
Understanding Electricity Prices: The Layers Behind the Price
When people talk about electricity prices, they often mean very different things. To make the debate more factual, it helps to break the price down into its layers.
-
China: The Carbon Giant Turning Green
While China is often headlined as the world’s largest emitter, there are some aspects that actually make me somewhat optimistic about their role.
-
Prosperity Without Pollution: The End of the Carbon-GDP Link
Historically, economic growth and CO₂ emissions were locked together: to get richer, you had to burn more. But the latest data shows this “iron law” is finally breaking.
-
The Climate Blame Game: Which Metric Actually Matters?
When judging a country’s climate impact, the “villain” changes depending on your metric.
-
Easter Stress Test for the German Power Grid
Good morning from Germany where the Easter Bunny is bringing another challenge for our energy grid! ☀️🔋
-
The Solar Paradox: Why the Equator Isn't the King of Solar
You might assume the Equator—where the sun hits most directly—would be the ultimate solar energy goldmine. But if you look at the Global Solar Atlas, the real “sweet spots” (the deep reds) aren’t at 0° latitude. Instead, they are concentrated in the subtropics (15°N to 35°N and 15°S to 35°S).
-
Three Takeaways from the OMR Podcast with E.ON's CEO
Last week, I listened to the OMR podcast featuring our CEO at E.ON, Leo Birnbaum. Among many interesting topics, let me comment on three of them:
-
Nuclear Power by the Numbers
The nuclear energy debate comes up from time to time, but my impression is that it’s rarely discussed based on numbers, but rather based on emotions. That isn’t very helpful if you actually want to understand the realities of the topic. So, I went directly to the IEA and the IAEA PRIS database to look up the hard data for myself.
-
Solar Potential: Windhoek vs. Berlin
In terms of LCOE (levelized costs of energy), solar is already the cheapest power source in Germany. But its potential is even higher in other regions. When you look at the raw solar fuel—the Direct Normal Irradiation (DNI)—available globally, the geographical advantages become undeniable. Comparing GlobalSolarAtlas.info data puts this into stark perspective: an identical solar panel in Windhoek, Namibia, has almost 3 times the raw energy to work with compared to one in Berlin, Germany (2,813 kWh/m² annually vs. 992 kWh/m²). Crucially, Windhoek also has much less seasonality. Its worst winter month for solar irradiation still significantly outperforms Berlin’s absolute best summer month, highlighting exactly why northern climates require massive seasonal storage to bridge their winter energy gaps.
-
The Blind Spot of LCOE: Solar Cannibalization
While the Levelized Cost of Electricity (LCOE)—the lifetime break-even cost per unit of energy—proves utility-scale solar is incredibly cheap to build, this metric has a major blind spot: it completely ignores when the electricity is generated. In reality, a kilowatt-hour is only truly valuable if it is produced at the exact moment the grid actually needs it.
-
Solar's learning curve: 20% cheaper with every doubling
The dramatic decline in solar costs is a powerful reminder of what is possible when technology scales globally. As the first chart from Our World in Data shows, for every doubling of cumulative capacity, module prices dropped by about 20%—moving from over $100/Watt in 1975 to just cents today.
-
German Electricity Prices: Residual Load and the Merit Order
German electricity prices are a story of two different timescales. On a daily timescale they are primarily driven by the Residual Load—the “gap” that remains after subtracting volatile renewable generation (wind and solar) from total demand. As seen in the first chart, when renewables are high, the residual load drops, often pushing prices toward zero. Conversely, when the wind isn’t blowing and the sun isn’t shining, we must rely on conventional power plants to fill the deficit, causing immediate price spikes to balance the grid.
-
Energy Efficiency in a Nutshell
When we talk about building renewable energy, we have to understand the difference between what we build and what we actually use. Here is how the math breaks down:
-
Earth's Axial Tilt and the Seasons: How 23.4° Drives Everything
The Earth’s axis is tilted by 23.4° — and that single fact drives everything we call “seasons.” These three animations show the chain of consequences:
-
Measuring What Matters: A Guide to Evaluating RAG Systems
From embedding-level sanity checks to end-to-end faithfulness metrics - here's how to know your RAG is working. -
RAG Unpacked: A Deep Dive into Core Components and Customization Strategies
This article dives into the critical choices and trade-offs at every stage of the RAG pipeline. -
Grounding LLMs: An Introduction to Retrieval Augmented Generation (RAG)
Learn how RAG connects LLMs to external knowledge, leading to more accurate and trustworthy AI answers. -
Der EU AI Act kommt
Was Unternehmen jetzt über Europas neue KI-Regeln wissen müssen. -
Async Python
This post breaks down asynchronous Python, offering a straightforward example to illustrate the sequence of execution. -
Navigating the Landscape of Retail Customer Support
A Foundation for AI-Driven Transformation. -
Einführung in Agentic Systems: Wie KI lernt zu denken, zu planen und Werkzeuge einzusetzen
Wie Agentic Systems durch schrittweises Denken, adaptive Planung und die Orchestrierung von Werkzeugen komplexe, mehrstufige Aufgaben lösen können. -
Recommender Systems in Financial Services
Recommender systems often begin with matrix factorization, but achieve greater nuance by adding explicit user and item features into more complex predictive models. -
XGBoost Deep Dive: From Decision Trees to Gradient Boosted Power
This post explains XGBoost's core recipe: sequential trees fitting loss gradients, boosted by stochasticity and a learning rate. -
Germany's Electric Grid: Keeping the Lights On Amidst the Green Shift
While renewables surge, Germany's grid data shows the ongoing need for conventional power to balance variability and keep the lights reliably on. -
KI-Risiken aktiv managen: Strategien für den sicheren Einsatz
Zur effektiven Minderung von KI-Risiken braucht es Datenqualität, robuste Systemarchitekturen und klare Verantwortlichkeiten. -
Strategic AI Implementation: How Companies Can Successfully Integrate AI – Now Available as Podcasts
What companies need to know about AI – and how they can integrate technology, people, and processes to achieve real impact. -
Künstliche Intelligenz mit Plan: So gelingt die Umsetzung im Unternehmen
Was Unternehmen über KI wissen müssen – und wie sie Technologie, Menschen und Prozesse zusammenbringen, um echte Wirkung zu erzielen. -
Semantic image segmentation with vector labels
A tutorial on approaches to deal with vector data labels in semantic image segmentation tasks. -
Image clipping with CRS reprojections
The notebook shows some of the intricacies of clipping satellite images when working with multiple CRSs. -
Open access geospatial data science books and courses
A list of open access geospatial data science books and courses that I found helpful -
Coordinate reference systems
Comparison of world maps with different coordinate reference systems -
Building a data science product with python: Transcription of audio conversations
Link to a video series that shows how to automatically transcribe mp3/wav podcasts with python using whisper and pyannote. -
ML in practice
Video of a presentation that I gave at KIT Founders Club about my experiences with ML / AI models in practice. -
Tools that make your life easier
A list of tools, machines and services that I use to make my life easier. -
Satellite image resolutions
A comparison of satellite image resolutions of different datasets. -
Asset management and automation
How much automation in asset management is the right amount? Should the computer make all the investment decisions itself? Where is human expertise still needed? Transcript of a podcast interview that I gave some years ago. -
Stable Diffusion: three ways to interact with the model (for free)
Stable diffusion is an open source text-to-image AI model. The model comes with multiple different access points that a user can use to generate images. In this post we describe three such ways and point to resources where you can play around with the model for free. -
Zonal statistics and US land coverage
Tutorial on computing zonal statistics with application to US land cover data -
Landsat 8 visualizations
Tutorial on Landsat satellite data: frequency of images and ways to visualize images without clouds -
ECB Statistical Data Warehouse
Tutorial on how to query ECB data and metadata with Python. -
BlackRock Advantage Series
The Advantage Series is combining technology, big data analysis and responsible investing. Using new and innovative data sources creates insights into financial markets and individual companies. Does this mean that the future of investing has arrived? -
The Carpentries
Get your coding and data science skills pumped up with workshops and publicly available lesson material of the Carpentries. Amongst others, learn how to use Unix Shell, Git, scientific computing languages (Python, R, Matlab), SQL, Make and Webscraping. -
Atlas of Sustainable Development Goals 2018
The Atlas of Sustainable Development Goals is a collection of socio-economic statistics and visualizations that measure the progress achieved towards the Sustainable Development Goals that are targeted by 2030.