Most of what I build lives in the Python ecosystem — data pipelines, machine learning, and the code that ties them together. The idea I keep returning to is treating the data model and the machine learning model as two halves of one design: each shaped by the other rather than bolted together after the fact.
In practice that runs from TensorFlow model code down to the pipelines that feed it. At Common Crawl it means turning terabytes of crawl data into an answer to a deceptively simple question — how much of the web does a crawl actually see?
An independent, read-only census of the agentic web. Every public registry that lists AI agents, MCP servers or agent catalogs is harvested in full, and every remote endpoint those listings point at is probed with the smallest real protocol handshake that proves it is what it claims to be. Registries publish what operators submit; none of them measure whether the entries answer. The time series is the product.
agentprobe.orgProcessing and analyzing 80-100TB of web data as part of my PhD research. Building scalable data pipelines with Scala and Apache Spark.
german-academic-web.deAI automation consulting for workflows. Most business processes are already a conversation between people, systems and documents that nobody designed. SwiftWerk maps that conversation and rebuilds it as an agent interaction architecture: which steps an agent owns, which tools and data it may touch, where agents hand work to each other, and where a human signs off. The goal is not one agent per task but fewer hand-offs, shorter cycle times and a process that can be measured end to end.
Architecture through optimization. A spiral staircase is a design problem fenced in by norms: DIN 18065 fixes ranges for rise, going, headroom and the walking line. Instead of checking a finished design against the standard, each rule enters the optimization as an energy term, a penalty that grows as the design strays from it. The staircase then settles into a low-energy shape, and the norms do the shaping rather than the policing.
Market research on calls for tenders. Each RFP on its own says little; read together, they show a market moving: what buyers ask for, how requirements and budgets shift, who bids and who wins. The aim is to turn that stream into value for both sides: tender writers can see how comparable calls were framed and answered, and applicants can see where demand is heading before the next call appears.