Eight years of commits across a startup, research, and enterprise AI consulting. The data below covers everything tracked in version control since October 2017.
My AI work is not a recent pivot. Data science tooling enters the commit history in 2018, ML frameworks from 2019, and production AI infrastructure from 2024. The stack covers research (PyTorch, Weights & Biases, HuggingFace Transformers, Optuna, SPARQL), production ML (LlamaIndex, Ray, Milvus, Cohere), and data science (Jupyter, NumPy, pandas, scikit-learn). I train models, track experiments, build retrieval pipelines, and run distributed inference. I also call APIs when that's the right tool, but that's not where the work ends.
I always have something running outside of work. During the Polygon years, that meant university research and open-source contributions alongside a more-than-full-time startup. During the bbv years, it meant ConceptFormer, the political bias study, 27 podcast episodes, and this website. The personal project heatmap has no gaps longer than a few months since 2017. I build things because I'm curious, not because someone is paying me to.
I write code in focused blocks during business hours, mostly between 9am and 7pm, with a peak around 2pm. Weekends exist in the data but stay quiet. The longest streak is 13 days. I don't romanticize late nights.
The more telling metric is the ratio between pull requests opened and pull requests reviewed: 580 opened, 901 reviewed. That ratio shifted over time. At PolygonSoftware I wrote most of the code myself. At bbv, my output is primarily reviews, architecture decisions, and issues that define work for others. In 2025 alone: 331 reviews, 202 issues opened, 131 PRs. I focus on making the team more productive rather than shipping individual features.
A model that works in a notebook is not a product. The gap between the two is infrastructure: APIs, containerization, cloud deployment, persistence, caching, messaging, orchestration, observability. I've built that layer for the Swiss AI Hub and multiple client deployments. Every component independently deployable, observable, and replaceable. No single tool dominates the commit distribution because a production platform can't afford single points of failure in its tooling either.
580,000 lines of JavaScript represent the PolygonSoftware years: 62 projects, 25 clients, five industries. Vue was the workhorse, React used where clients required it, the stack evolving as the company scaled. This was full-stack product engineering under time pressure and budget constraints, choosing frameworks based on client needs rather than personal preference. The frontend work continues at bbv through the AI Hub's interface, but at lower volume. Web development went from being the entire job to being one capability among several.
The lines-of-code chart compresses the whole story into one image. Java on the left is university. The JavaScript wall in the middle is a startup being built at full speed. The Python curve on the right is what came after: research, AI infrastructure, and a deliberate shift toward the work I want to do for the next decade.
Joel's AI stack spans research (PyTorch, Weights & Biases, HuggingFace Transformers, Optuna, SPARQL), production ML (LlamaIndex, Ray, Milvus, Cohere), and data science (Jupyter, NumPy, pandas, scikit-learn). He trains models, tracks experiments, builds retrieval pipelines, and runs distributed inference.
Joel builds production infrastructure using FastAPI for APIs, Docker for containerization, AWS and Azure for cloud deployment, Redis for caching, NATS for messaging, MongoDB and Milvus for persistence, Dagster for orchestration, and OpenTelemetry for observability. Every component is independently deployable, observable, and replaceable.
Joel has written 580,000 lines of JavaScript across 62 projects for 25 clients in five industries during the PolygonSoftware years. Primary stack was Vue.js with Nuxt, TailwindCSS, and various chart libraries. React was used where clients required it. This was full-stack product engineering under time and budget constraints.
Joel has opened 580 pull requests and reviewed 901, with the ratio shifting over time. At PolygonSoftware he wrote most code himself. At bbv, his output is primarily reviews, architecture decisions, and issues defining work for others. In 2025 alone: 331 reviews, 202 issues opened, 131 PRs. He sees himself as making teams more productive.
Joel writes code in focused blocks during business hours, mostly between 9am and 7pm, with peak productivity around 2pm. Weekends exist in the data but stay quiet. His longest streak is 13 days. He doesn't romanticize late nights, preferring sustainable work patterns.
.
Copyright 2026 - Joel P. Barmettler