In a time when AI is changing so much around us, one thing becomes even more important: people who support each other have a real advantage. Communities that share knowledge, exchange ideas, and help each other grow will always move further, faster.
Beyond the talks and the program, PyCon DE is all about people: the conversations in the hallways, the familiar faces, and the new connections that start here.
Beyond the talks and the program, PyCon DE is all about people: the conversations in the hallways, the familiar faces, and the new connections that start here.
Media is too big
VIEW IN TELEGRAM
What stayed with us from PyCon DE 2026
Three messages came through the loudest:
— Open source is the foundation, not the add-on.
The keynotes — Sebastian Raschka, Gabriela Bogk, Hilde Kühne, Aaron Glenn — kept returning to the same point: the AI systems we can trust are the ones we can inspect, run, and adapt ourselves.
— Software engineering matters more, not less, in the age of LLMs.
Rigor, testing, maintainability — that’s the difference between a demo and a system that actually holds up in production.
— Data engineering is where most real problems get solved.
And still the part of the stack that doesn’t get enough airtime in the public AI conversation.
The Open Source Business panel (Probabl, QuantStack, Explosion) and the Startup Lounge with hessian.AI made the economic side tangible: European open source companies are building, hiring, and shipping — right now.
#PyConDE #PyData #Python #AI #OpenSource #DataEngineering
Three messages came through the loudest:
— Open source is the foundation, not the add-on.
The keynotes — Sebastian Raschka, Gabriela Bogk, Hilde Kühne, Aaron Glenn — kept returning to the same point: the AI systems we can trust are the ones we can inspect, run, and adapt ourselves.
— Software engineering matters more, not less, in the age of LLMs.
Rigor, testing, maintainability — that’s the difference between a demo and a system that actually holds up in production.
— Data engineering is where most real problems get solved.
And still the part of the stack that doesn’t get enough airtime in the public AI conversation.
The Open Source Business panel (Probabl, QuantStack, Explosion) and the Startup Lounge with hessian.AI made the economic side tangible: European open source companies are building, hiring, and shipping — right now.
#PyConDE #PyData #Python #AI #OpenSource #DataEngineering
👍2
📅 Save the date: April 20–22, 2027 — Heidelberg
PyCon DE has already started preparing for the next edition.
Join us for a few days of talks, panels, and discussions on Python, AI, data, and the future of software development — from hands-on engineering practices to real-world applications of machine learning and LLMs.
Beyond the program, it’s also about meeting the community: reconnecting with people you know, starting new conversations, and exchanging ideas with developers, data scientists, and teams building real systems.
We’re looking forward to bringing everyone together again — in a new city, with new topics and new perspectives.
See you in Heidelberg 🐍💻✨
PyCon DE has already started preparing for the next edition.
Join us for a few days of talks, panels, and discussions on Python, AI, data, and the future of software development — from hands-on engineering practices to real-world applications of machine learning and LLMs.
Beyond the program, it’s also about meeting the community: reconnecting with people you know, starting new conversations, and exchanging ideas with developers, data scientists, and teams building real systems.
We’re looking forward to bringing everyone together again — in a new city, with new topics and new perspectives.
See you in Heidelberg 🐍💻✨
🔥4
"Honey, I vibe coded some crypto."
Our keynote from last conference, Gabriela Bogk, started with a real story. A team vibe-coded a cryptographic implementation in a hackathon week, then asked their CISO whether it was secure. It wasn't. So she pointed AI at the same code, told it exactly what to look for, and it found the vulnerabilities, round after round.
Real leverage, then. The catch is that you cannot trust its own proof of work: integration tests written against a complete mock, a test that passes by printing "success". Everything green, nothing verified.
Which is why the checks belong outside the AI. Guardrails in code, not in prompts, because a prompt can always be talked around. And if a problem is hard to solve but easy to verify, write the code that does the verifying.
Every team is deciding right now how much to let AI do. Bring yours to PyCon DE and let's work it out together.
Full keynote: https://pycon.de/archive/2026/talks/honey-i-vibe-coded-some-crypto-security-in-the-age-of-llms/
Our keynote from last conference, Gabriela Bogk, started with a real story. A team vibe-coded a cryptographic implementation in a hackathon week, then asked their CISO whether it was secure. It wasn't. So she pointed AI at the same code, told it exactly what to look for, and it found the vulnerabilities, round after round.
Real leverage, then. The catch is that you cannot trust its own proof of work: integration tests written against a complete mock, a test that passes by printing "success". Everything green, nothing verified.
Which is why the checks belong outside the AI. Guardrails in code, not in prompts, because a prompt can always be talked around. And if a problem is hard to solve but easy to verify, write the code that does the verifying.
Every team is deciding right now how much to let AI do. Bring yours to PyCon DE and let's work it out together.
Full keynote: https://pycon.de/archive/2026/talks/honey-i-vibe-coded-some-crypto-security-in-the-age-of-llms/
🔥4
"I want to clean my car. I live 200 metres from a car wash. Should I walk or drive?"
It came back with pros, cons and environmental arguments, and missed that you need the car.
Silke Horn showed that at our last conference, in a talk on optimization: problems where you need the best answer and a proof that it is the best. How the NFL fits 272 matches into a season.
Then she asked ChatGPT to solve her assignment problem. It did, by recognising the structure and calling a solver. Every model she tried did the same, and forced not to, answered suboptimally.
An LLM is good at spotting what kind of problem you have. It gives no guarantees and cannot say why one option beat another. A solver can.
Knowing which problems need a proof is the takeaway. PyCon DE is full of talks like it. We get things done.
🎥 Full talk: https://pycon.de/archive/2026/talks/from-hard-problems-to-proven-solutions-solving-decision-problems-with-gurobi/
📍 PyCon DE 2027 · Heidelberg · 19–23 April
#PyConDE #Optimization
It came back with pros, cons and environmental arguments, and missed that you need the car.
Silke Horn showed that at our last conference, in a talk on optimization: problems where you need the best answer and a proof that it is the best. How the NFL fits 272 matches into a season.
Then she asked ChatGPT to solve her assignment problem. It did, by recognising the structure and calling a solver. Every model she tried did the same, and forced not to, answered suboptimally.
An LLM is good at spotting what kind of problem you have. It gives no guarantees and cannot say why one option beat another. A solver can.
Knowing which problems need a proof is the takeaway. PyCon DE is full of talks like it. We get things done.
🎥 Full talk: https://pycon.de/archive/2026/talks/from-hard-problems-to-proven-solutions-solving-decision-problems-with-gurobi/
📍 PyCon DE 2027 · Heidelberg · 19–23 April
#PyConDE #Optimization
🔥1
The team that built the last PyCon is coming back for Heidelberg.
As we started planning PyCon DE 2027, almost everyone from the last team signed up to do it all over again. That kind of continuity is a gift, and it is the clearest sign we know that this community is worth returning to.
It also means there are no open seats on the organising team right now, so this year we are not running a Call for Committees. If you were hoping to join the crew, this is not a no forever. Teams change from year to year, and we would love to have you when a spot opens up.
And there are still two great ways to help shape the conference. The Call for Reviewers lets you read, rate, and help select the talks that make up the program, one of the most direct ways to influence what happens on stage. The Call for Volunteers puts you on the team that makes the event run on-site, and it is one of the best ways to get to know the community from the inside.
We will open both a little later on. Keep an eye on our website and channel.
As we started planning PyCon DE 2027, almost everyone from the last team signed up to do it all over again. That kind of continuity is a gift, and it is the clearest sign we know that this community is worth returning to.
It also means there are no open seats on the organising team right now, so this year we are not running a Call for Committees. If you were hoping to join the crew, this is not a no forever. Teams change from year to year, and we would love to have you when a spot opens up.
And there are still two great ways to help shape the conference. The Call for Reviewers lets you read, rate, and help select the talks that make up the program, one of the most direct ways to influence what happens on stage. The Call for Volunteers puts you on the team that makes the event run on-site, and it is one of the best ways to get to know the community from the inside.
We will open both a little later on. Keep an eye on our website and channel.
❤5
Building a Streamlit dashboard took days. Deploying it took weeks.
That was Bernhard Schäfer's experience as a consultant. At our last conference he and Nicolas Renkamp showed Merck's answer. Name the app, pick a template, link it to a use case, click launch. Three minutes later it runs, with single sign-on, CI/CD and governed data access. Close to 700 apps are live, at around 21 euro each a month.
Nicolas named the governance model, not the speed, as the key factor. Every app is tied to a use case, and the token is scoped to that use case's data, so an app cannot reach everything its user can see. That is what lets it use production data from day one, without a separate code review.
Not open source, but Nicolas said the talk gave enough detail to build something like it. We get things done.
🎥 Full talk: https://pycon.de/archive/2026/talks/empowering-data-scientists-with-zero-platform-friction-deploying-streamlit-friends-in-3-minutes/
📍 PyCon DE 2027 · Heidelberg · 19–24 April
#PyConDE #Python
That was Bernhard Schäfer's experience as a consultant. At our last conference he and Nicolas Renkamp showed Merck's answer. Name the app, pick a template, link it to a use case, click launch. Three minutes later it runs, with single sign-on, CI/CD and governed data access. Close to 700 apps are live, at around 21 euro each a month.
Nicolas named the governance model, not the speed, as the key factor. Every app is tied to a use case, and the token is scoped to that use case's data, so an app cannot reach everything its user can see. That is what lets it use production data from day one, without a separate code review.
Not open source, but Nicolas said the talk gave enough detail to build something like it. We get things done.
🎥 Full talk: https://pycon.de/archive/2026/talks/empowering-data-scientists-with-zero-platform-friction-deploying-streamlit-friends-in-3-minutes/
📍 PyCon DE 2027 · Heidelberg · 19–24 April
#PyConDE #Python
🔥2
190 Python repositories, one shared template, almost no manual merge-conflict cleanup.
At our last conference, Yannik Tausch of QuantCo walked through how they keep that many repositories in sync. The core move is Copier instead of Cookiecutter: it remembers the answers from when a project was generated, so updating later is a three-way git merge between the current code, the old template and the new one. Customizations survive.
The part worth remembering is how they roll out something riskier, like a new security linter. A normal update would break CI everywhere at once. Instead it ships as a template question defaulting to false, and a bot flips it to true per repository once ready.
Yannik also contributes to conda-forge — an open-source maintainer our program keeps bringing into the room.
🎥 Full talk: https://pycon.de/archive/2026/talks/kickstart-coding-at-scale-how-project-template-automation-unlocks-developer-productivity/
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
At our last conference, Yannik Tausch of QuantCo walked through how they keep that many repositories in sync. The core move is Copier instead of Cookiecutter: it remembers the answers from when a project was generated, so updating later is a three-way git merge between the current code, the old template and the new one. Customizations survive.
The part worth remembering is how they roll out something riskier, like a new security linter. A normal update would break CI everywhere at once. Instead it ships as a template question defaulting to false, and a bot flips it to true per repository once ready.
Yannik also contributes to conda-forge — an open-source maintainer our program keeps bringing into the room.
🎥 Full talk: https://pycon.de/archive/2026/talks/kickstart-coding-at-scale-how-project-template-automation-unlocks-developer-productivity/
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
🔥1
Splitting code into ports and adapters doesn't just make it easy to swap a SQL database for a NoSQL one — it also makes the code easier to hand off to an AI assistant.
At our last conference, Luke Gerstner of Rosenxt walked through how he uses hexagonal architecture (ports and adapters, from Alistair Cockburn in the early 2000s).
🎥 Full talk: https://pycon.de/archive/2026/talks/leveraging-hexagonal-architecture-when-building-applications/
The domain model sits at the center, ports define contracts as Python abstract base classes, and adapters hold the database code — swap SQL for NoSQL and only the adapter changes. Luke's strongest case for it is AI: separated from the plumbing, an assistant like Claude can generate a lot of that plumbing on its own. He showed this on a sensor-installation service at Rosenxt.
The Q&A ran long afterward, from multi-agent systems to Robert Martin's clean architecture — worth being in the room for.
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
At our last conference, Luke Gerstner of Rosenxt walked through how he uses hexagonal architecture (ports and adapters, from Alistair Cockburn in the early 2000s).
🎥 Full talk: https://pycon.de/archive/2026/talks/leveraging-hexagonal-architecture-when-building-applications/
The domain model sits at the center, ports define contracts as Python abstract base classes, and adapters hold the database code — swap SQL for NoSQL and only the adapter changes. Luke's strongest case for it is AI: separated from the plumbing, an assistant like Claude can generate a lot of that plumbing on its own. He showed this on a sensor-installation service at Rosenxt.
The Q&A ran long afterward, from multi-agent systems to Robert Martin's clean architecture — worth being in the room for.
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
🔥1
A greedy Python script took 38 color changes to paint a car sequence in alternating black-or-white base coats. The mathematically proven optimal: 23.
At our last conference, Justine Broihan of GAMS Software GmbH put the room through that same decision live, on a smaller scale — guess the fewest color changes, no reordering allowed. Most answers came from a greedy rule: keep one color until forced to switch.
🎥 Full talk: https://pycon.de/archive/2026/talks/the-art-of-the-optimal-a-pythonic-approach-to-complex-decision-making/
Mathematical optimization describes the problem itself — variables, constraints — for a solver that proves optimality. GAMSpy, Justine's team's Python library, interfaces with 36 solvers. Colleague Muhammet Soyturk, its main developer, paired it with ML: a PyTorch model embedded as a constraint, keeping defects under 5%.
In the room, you'd have shouted your own guess before Justine showed how far off most were.
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
At our last conference, Justine Broihan of GAMS Software GmbH put the room through that same decision live, on a smaller scale — guess the fewest color changes, no reordering allowed. Most answers came from a greedy rule: keep one color until forced to switch.
🎥 Full talk: https://pycon.de/archive/2026/talks/the-art-of-the-optimal-a-pythonic-approach-to-complex-decision-making/
Mathematical optimization describes the problem itself — variables, constraints — for a solver that proves optimality. GAMSpy, Justine's team's Python library, interfaces with 36 solvers. Colleague Muhammet Soyturk, its main developer, paired it with ML: a PyTorch model embedded as a constraint, keeping defects under 5%.
In the room, you'd have shouted your own guess before Justine showed how far off most were.
📍 Join us at PyCon DE 2027, Heidelberg · 19–24 April
#PyConDE #Python
🔥2