About me

I'm Joel. I research and build AI systems. I've also run a company, which informs how I approach everything else.

The distance between what AI can do in a paper and what it can do in production is where I work. Bridging that gap requires research and engineering, and I've ended up doing both.

Building and shipping

At 22, while studying computer science at the University of Zurich, I co-founded PolygonSoftware with two classmates. Over four years we grew from three founders to 18 engineers, delivered 62 projects for 25 clients across five industries, and learned what it takes to keep a software company alive.

I ran over 100 sales meetings and 80 pre-sales calls. I conducted 40+ coding interviews and hired 26 people. I made technical decisions that saved us and technical decisions that cost us. The company ended in 2023 through a failed merger. What survived: an instinct for what customers actually need and a respect for the gap between a working demo and a system that runs in production for paying users.

→ More on PolygonSoftware

Research

I finished my master's at UZH summa cum laude, with a thesis graded 6.0. The work became ConceptFormer, a method for grounding language models in knowledge graphs without converting structured data to text. It was published at the GLOW workshop at WWW'26.

The research question that drives me: how do we make language models trustworthy enough to deploy in contexts where being wrong has consequences? This pulls toward knowledge representation, factual grounding, and understanding where these systems fail. My study on political orientation in 21 language models tested them against Swiss voting recommendations. The instinct was the same: if we're going to deploy these systems, we should understand what they've absorbed and what they'll say when asked.

→ Research

AI infrastructure

Since 2023 I've worked as Senior AI Consultant at bbv Software Services, where I architect the AI-Hub platform: the infrastructure layer that lets enterprise clients deploy language models with retrieval, tool use, and observability. Fifteen clients, twelve projects.

This is not wrapper code around OpenAI. The stack includes vector databases (Milvus), distributed compute (Ray), data orchestration (Dagster), message streaming (NATS), and observability throughout (OpenTelemetry). Building systems at this layer has taught me where LLM applications actually break. It's rarely the model.

→ More on bbv

Communicating

I host Attention Heads, a German-language podcast on AI with 27 episodes covering technical deep-dives, industry analysis, and the social implications of these systems. I've given five conference talks, including panels at the Swiss AI Impact Forum, and hosted six webinars on topics from ChatGPT internals to AI agents in enterprise.

Technical people who can explain are useful. I've found that forcing myself to communicate research and engineering clearly to non-specialists sharpens my own understanding.

→ Podcast · Appearances

What I'm looking for

I want to work on AI systems where research quality and production reliability both matter. I'm looking for roles that take the gap between papers and deployment seriously, whether in an industry research lab or a PhD program with strong engineering culture.

For the raw data behind any of this, see Skills.