Building end-to-end ML systems can be challenging, especially when it's part of a complex user experience in a health app. This is a story of how a user-focused mindset and product thinking helped Flo  Health to build the best ML powered cycle predictions on the market.

Flo Health is the no. 1 female health app in the world. One of the core ML tasks for us is the "cycle start date" prediction problem. During the talk I'll share the difficulties we faced, the learnings we gathered and how important and helpful it might be to always think about the end users of the system you build.

Ivan Klimuk

User-Centric Machine Learning

Lightning talks are designed to be short presentations between five and ten minutes long, but are usually capped at five minutes. Most conferences will allot a segment of roughly 30 to 90 minutes long to speakers. Talks are arranged one after the other during the sessions. The talks are usually given at conferences in order for the event to have many speakers discuss a multitude of topics. The conferences are held in order for individuals to be able to share their ideas and concepts to people who have experience in the specific field.

Lightning Talks 1

This session is about a simple and flexible ML microservices engine called Skipper. It is open source solution and it runs on Python. During this technical session, you will learn how to run FastAPI endpoints on Kubernetes and enable microservices communication through RabbitMQ with Python API. Besides technical info, the author will share why he was inspired to build this solution and how he participates in Python community.

Andrej Baranovskij

MLOps with FastAPI, RabbitMQ and Kubernetes using Skipper

Deep learning applications have been adopted widely across many industries. Radiology is no exception - it certainly can be considered a breakthrough technology. Healthcare institutions have put trust in Oxipit's Deep Learning technology to improve their Radiology workflow. We will discuss necessary steps for building Deep Learning solutions using Oxipit as an example.

Darius Barušauskas

Artificial Intelligence in Radiology - are we there yet?

Sharing and explaining the results of your analysis can be a lot easier and much more fun when you can create an animated story of the charts containing your insights. [ipyvizzu](https://github.com/vizzuhq/ipyvizzu) enables just that using a simple Python interface. 

In this workshop, one of the creators of ipyvizzu introduces this tool and helps the audience in taking the first steps in utilizing the power of animation in storytelling. After the workshop, the members will be able to build and present animated data stories on their own.

Outline of the workshop:
- Introduction to Vizzu, the open-source C++/JavaScript lib behind ipyvizzu
- Key concepts when using a generic chart building and morphing engine.
- The 4 key parts of the animated method - the central element of ipyvizzu
- Adding data with and without using pandas
- Chart configuration options
- Styling
- Animation options
- Q&A

Peter Vidos

ipyvizzu - a new, open-source charting tool to create animated data stories with Python in Jupyter

Machine learning (ML) models rarely make it to production. Often these projects start with intention ‘let’s make something cool’, but they get stuck in local Jupyter notebooks or Power Point presentations. What does it take to complete a data-driven product?
I will talk about the importance of the problem and it's context definition, will give examples of overcomplication and together we will go through the steps required to build a data-driven product.

Karolina Griciunė

Data-driven products: from zero to hero

The great learning experience is a stepping stone to becoming a professional in your field. All you need to have is good motivation and a proper toolset. Having first steps in the right direction is crucial for a successful career and we’re here to help future developers.


This talk will introduce you to the latest tool we all were looking for.

The IT sector is thriving, but it still can’t get enough students interested in it. First impressions are often the most important, so maybe the tools are at fault? How can you be interested in the field if you are overwhelmed with over-complicated syntax or a 3 centuries-old toolset at your first and most important steps? After some investigation and analysis, we created a new project, “Angis” - an online free-to-use learning platform based on the amazing beginner-friendly programming language Python. So after a successful project, the next step is to have a proper textbook for students in their native language.

Rokas Cvirka

Artūras Nikončukas

Python Handbook for schools

Ray tracing is a rendering method. It simulates the physical behavior of light and it allows to achieve generation of very realistic images.

In this talk I will show an old-school CPU based Python implementation. Python is relatively slow programming language so a ray tracer written in Python is also going to be slow. I will show several (some of them hacky) tricks to speed up CPU intense calculations.

Petras Zdanavičius

Implementing Ray Tracing in Python

Locking could be done in many different ways, some of them are quite heavy (slow), others can be dangerous (deadlocks), in this session we will go through some simple ways of locking a resource and profile the different approaches. 

We will go through the algorithm of RedLock which could be implemented with many others storage solutions, we will see pros and cons and if time allows also its implementation with multiple Redis instances. 

For the demo Django and Redis will be used and to profile the http calls locust will be used.

Redis can be much more than a cache, one of the things that you can easily do with it is have locks to synchronise your instances in a very lightweight manner

Anas El Amraoui

RedLock a smart lock for distributed systems

We propose a new method of testing that seems to work really well, but is not formalized in any way. It has been applied in 3 projects of which 2 were serious business (the other is a hackathon project).

No existing testing framework (python or otherwise) embraces this approach, but the core of the idea fits in 50 lines of code.

A generally accepted theory of "testing trophy" states that most of the tests we write should be integration tests. From this we conclude that the mantra "tests should be independent" that makes sense for unit testing, but has been expected from integration tests as well is actually wrong.

The main motivation to have independent tests is to avoid chaos and have a stable repeatable testing environment. We argue that such artificially imposed stability is a source of bugs and a great maintenance burden.

Instead we suggest to make test interdependence explicit and apply the idea that one tests output is the next tests input. For example test_update_user might depend on test_create_user instead of mocking the state of an already created user.

Pēteris Ratnieks

A new framework for testing

We're always using type hints in our code, but when working with microservices or when doing some integrations I started noticing that type hints alone are not enough to help us. In this presentation, I'll try to show my experience of how I moved and what benefits I got from using dict type hints to TypedDict, then to Dataclasses, and lastly to Pydantic. At the end of the day, it helped me to save a lot of time not only when implementing features, but also with the maintenance of already implemented parts of code.

Justinas Kuizinas

TypedDict, Dataclasses and Pydantic in action

Websockets protocol has provided bi-directional communication capabilities for web developers. However, the scalability of Websockets is not as simple as stateless REST API. In this presentation I will define a basic architecture using Websockets and try to scale it. The measured results should provide a better understanding what throughput capabilities of Websockets are and whether this protocol might be suited for your needs.

Albertas Gimbutas

Scaling Websockets

Recently, a paper titled [Trojan Source](https://trojansource.codes/) is published to demonstrate how a visibly valid contribution can contain malicious code by exporting the Unicode control characters. Some of these have been tested on Python and it works. How can it happen? Shall the Python and open-source communities be concerned?

### Background:

After researchers at the University of Cambridge published [a paper](https://trojansource.codes/trojan-source.pdf) about a malicious attack named Trojan Source, which exploited the fact that some program interpreters, like CPython, can handle Unicode. This has caused concerns in the open-source community about the malicious contribution that looks totally legitimate in human eyes but contains invisible attacks. As a member of the Python community, we should all be aware of that and understand how we can prevent this attack to happen.

### About this talk:

In this talk, Cheuk will decode the finding in this paper to a level that can be understood by everyone. She will start with a joke example of how you can mess up someone by using Unicode. She will then explain what is Unicode and why it causes trouble. Afterwards, she will explain the [Python examples](https://github.com/nickboucher/trojan-source/tree/main/Python) in the paper and why it can be dangerous. Lastly, she will open up a discussion on how we should defend ourselves from those attacks and what we can do as a community.

### Outline (30 mins talk):

**5 minutes - Introduction, the opening of the talk**

In this session, Cheuk will ask audiences to debug a code snippet that looks absolutely fine but will not work as code. She will explain that this is the same concept used in Trojan Source.

**10 mins - What is Unicode**

In this session, Cheuk will give an introduction about what is Unicode, what it is to a computer and why we need Unicode in computers. She will also explain how the benefit of having Unicode can also be a downfall to making us vulnerable to the Trojan Source attack.

**10 mins - How Trojan Source works in Python**

In this session, Cheuk will show a few examples using the Trojan Source in legitimate Python code. She will point out how the attack is hiding in the source code and in what cases it can be dangerous.

**5 mins - How to protect ourselves**

In this session, Cheuk will open the discussion and make a few suggestions of how we can protect ourselves as a community. This will lead to the Q&A session where the audience can weigh in on their own thought.

### Target audiences

From those who are curious to maintainers of open-source libraries. This is the knowledge we should all know and be aware of. Cheuk will explain in a way that expects no prior knowledge is needed.

### What will audiences learn

About Trojan Source attacks and how it works. They may also learn about how interpreters, especially Python interpreters, work with Unicode. Plus, they may have increased awareness about security in the open-source world.

Cheuk Ting Ho

Trojan Source Code - Can we trust open-source anymore?

Almost every time in a data analysis, you will inevitable find the presence of unexpected or weird values in your data. The majority of statistical and machine learning algorithms will fail to converge  or generalize with dirty data, therefore It is critical for the analyst to know how to identify and remove outliers in the data. 

In this talk, I will show you the most common techniques to eliminate outliers in the data using Python, and will give you useful tips on how to spot them.

- What are outliers
- How outliers can affect your analysis
- Difference between noise and outliers
- Visualizations tools
- Statistical methods
- Automatic methods
- Removing outliers
- Conclusions and Recommendations

Sara Iris Garcia

Detecting and removing outliers in your data

I am biased. I cannot stop seeing complex systems.  Complex systems are characterised by many components that interact in multiple ways among each other and with their environment. That lens of complexity science enables to see no limits in scaling experimentation at Vinted. Experimentation has many components to it. From low level - what sort of hashing algorithm you use for randomisation? To medium level - how do you scale metrics computation? To what statistical methods you use to ensure meeting high level business objectives? And you know what is the common thread among all these layers at Vinted - it is Python!

Jev Gamper

Up and Down the Ladder of Experimentation

In todays modern world it is no longer enough to know the programming language. It is even no longer enough to know a framework. Devops practices become more and more engrained into developers day-to-day activities. In this talk we will show you how to utilise AWS CDK to write python code that manages infrastructure.

Laimonas Sutkus

AWS CDK with Python

The pandas library is one of the key factors that enabled the growth of Python in the Data Science industry and continues to help data scientists thrive almost 15 years after its creation. Because of this success, nowadays there are several open-source projects that claim to improve pandas in various ways.

In this talk we will go over some of the most widely used dataframe Python libraries beyond pandas, clarify the relationship between them, compare them in terms of project scope and proximity to the original pandas API, and offer advice on when to use each of them.

The pandas library is one of the key factors that enabled the growth of Python in the Data Science industry and continues to help data scientists thrive almost 15 years after its creation. Because of this success, nowadays there are several open-source projects that claim to improve pandas in various ways, either by bringing it to a distributed computing setting (Dask), accelerating its performance with minimal changes (Modin), or offering slightly different API that solves some of its shortcomings (Polars).

The outline of the talk goes as follows:

1. Short introduction to the importance of pandas, and brief recollection of its main pain points (5 minutes)
2. Enumeration of some alternatives, description of our classification (pandas-like vs bespoke, single-node vs distributed) (5 minutes)
3. Presentation of the libraries using brief code snippets, visualization of the dependency relationships between them (20 minutes)
4. Recommendations and conclusions (5 minutes)

After the talk, you will have more information on how some of the modern alternatives to pandas fit onto the ecosystem, understand which ones provide the easiest migration path for an existing codebase, and be more prepared to judge which one to use for your next project. Prior exposure to pandas will help make the most of the presentation.

Juan Luis Cano Rodríguez

Beyond pandas: The great Python dataframe showdown

The situation of the past two years has served as a catalyst for further digitalisation of all sorts of administrative processes all over the world. In virtually all of these processes, public-key infrastructure (PKI) plays an important role.

In this talk, you'll get a crash course on what the modern PKI landscape looks like, and on the role that Python can play in making digital signing workflows accessible to a wide range of users in a responsible manner.

Most developers are familiar with PKI in the form of SSL certificates, an important component in the system that makes HTTPS secure.
However, PKI is much more than that: for example, in a digital workflow, digital signatures backed by public-key cryptography are a useful tool to ensure that all parties stay accountable to one another.

Of course, the basic technologies behind PKI are by no means new; governments and companies have been working with them for decades now. But the landscape is changing, and many "ordinary people" find themselves having to use PKI in their daily lives more frequently---sometimes without being aware of it!

As companies rush to carve out their slice of this growing market, I think it's important to raise awareness in developer communities: demystifying PKI is a good thing for us _and_ for our users.


I've written lots of code that interfaces with PKI, both in my own projects and on the job. I believe there's a lot of potential in the Python community to help provide tooling to further democratise PKI and digital signing. Hopefully this talk will contribute to that goal!

Matthias Valvekens

Bringing digital signatures and PKI to the masses with Python

Tips and tricks to properly mix async code with blocking code, using modern libraries, and taking advantage of the new concurrency features of the language while keeping compatibility with already existing code and libraries.

It will include the basics of what concurrency is, how to take advantage of it in the simplest use cases, and how to safely mix it with regular code.

You've heard async is great, it can improve concurrency performance, but you might have heard that you need to use only async code and use only async-compatible libraries... but the good news is, that's not entirely true.

So, how can you take advantage of these new async features in your current projects that already use blocking code and require some blocking libraries?

In this talk, I'll show you the basics of writing concurrent async code, and how to properly mix it with regular code without affecting performance (without blocking the event loop) by using thread workers underneath. All done in a simple and convenient way using modern libraries.

Sebastián Ramírez

Await for it: mixing async and blocking code

Type annotations are documentation. They're meant for humans to quickly figure out what arguments a given function accepts and what it returns. But can this information be utilized by Python to make code run faster? Yes and no!

In this talk, we'll cover a few methods of optimizing code using type annotations. We'll talk about why good human-readable types aren't the same as types that a machine would want to get to optimize. We'll look at real-world optimized Cython code from EdgeDB, and real-world code compiled with mypyc (the Black auto-formatter).

Łukasz Langa

Use typing to speed up your apps with mypyc

Given the vast amount of resources we have online on becoming a data scientist, it is only natural to feel overwhelmed and lost. Those who manage to start a career in the field, eventually spend majority of the time trying to find their dreamed (non-existent) position [Financial Times, 2017]. I am not here to give you extra unnecessary resources, I am here to showcase how you can leverage your interests to achieve a successful data science career. Bring passion, curiosity and a computer. You need no more!

Andrea Coifman

It's (not) about the data!

At [ZealiD](https://www.zealid.com/en/), we had to implement eMRTD RFID / NFC chip reading, parsing and cryptographic verification in Python. This proved to be an interesting excursion into ASN.1 data structures and handling crypto in Python. I'll walk you through this journey of discovery and will highlight Python's strengths and gotchas here.

We had to make sure we do this properly as it forms part of our qualified remote TRA process and qualified certificate issuance.

Combining ability to traverse and prototype `asn1crypto` library structures and later apply `mypy` proved useful. I'll show you why!

See Abstract. Apart from that, the context is: we had to implement [eMRTD](https://en.wikipedia.org/wiki/Biometric_passport) (travel/identity docs) [ICAO 9303](https://www.icao.int/publications/pages/publication.aspx?docnum=9303)-compatible RFID / NFC chip data reading, parsing and [cryptographic verification](https://en.wikipedia.org/wiki/Biometric_passport#Data_protection) in Python for the purposes of Passive Authentication.

We did not find any decent Python library to do this. All we found were snippets which make hacky openssl calls, ingest results, do ugly untyped bit wrangling and then do some more openssl calls. We ended up writing our own solution for this.

Kostas Jakeliūnas

Processing identity doc NFC chips: Pythonic way to handle parsing and cryptography

We decided to make a coding cool for the young generation … and created a modern gamelike online tool for kids to explore programming. In this talk we’ll talk about the bumpy road we had during this fantastic journey. Morally, the mission was about modernising the approach to teach coding. Technically, it threw us on a  neck breaking adventure from hacking a compiler to writing a high scalability SaaS platform to making Python wrappers around 2D JS game engine to writing serverless lambda functions on Azure and what not.

We’ll share the details about the educational goals we set ourselves as well as some juicy technical bits. 

Please, feel free to try the online learn to code tool Angis we’ve built during this journey (Note: we’ve hacked the Python into supporting Lithuanian, so it’s all Lithuanian by default, but the code editor supports English too).

About the speakers:

Gedas Lukšas - software developer who's also passionate about learning and teaching.

Mantas Urbonas - Mantas - Visma Lietuva CEO and one of the key people behind project Angis.

Gedas Lukšas

Mantas Urbonas

Hacking a Python Compiler, a Game Engine and Some Other Dark Magic to Build a Pythonic Online Learn to Code Tool for Kids

A few weeks ago, Py:Script was announced. This means that Python can now run in the browser, no transpiling or compiling of any kind needed.

This is great news but does it mean we will now be able to run well-known Python web frameworks in the browser? Or maybe we need to make schmol adjustments in order to make it work?

PyScript is new and somewhat revolutionary. One question that usually comes after that is "Can we run Django/Pyramid/\<plug you favorite framework here\> as a frontend framework?"

The short answer is *no*. Python web frameworks center a lot around the request/response cycle which doesn't happen in a frontend framework. There are also other issues, like file access. This doesn't mean that we cannot re-use existing tools that were developed for the backend and use them in a frontend framework.

This talk will look at what a usual frontend framework provides and whether Python and Py:Script can provide the same functionalities.

We will also take a look at a Proof-Of-Concept library that tries to provide those features.

Emma Delescolle

Py:Script: could it be used as a frontend framework?

Deploying a machine learning model to production the right way is not a trivial task, and involves many components. This talk aims first to walk the listener through the realm of MLOps by reviewing the typical challenges faced when deploying machine learning models, and then to flesh out a mature MLOps setup using the Databricks platform.

In recent years, the ability to develop complex machine learning models has immensely progressed. Together with such advances came the need to operationalize models in a mature way, a topic referred to as "MLOps". In March 2020, Google issued its famous 3 levels of MLOps maturity, from the most basic to advanced machine learning software development and deployment lifecycle. However nowadays many companies still struggle to deploy their machine learning models in a mature way.

The first part of the talk will detail what constitutes a mature MLOps and describe essential components such as the feature store, the model registry, the CI/CD pipeline, or the continuous model retraining. The second part of the talk will flesh out these concepts with a batch use case using the Databricks platform.

Philippe de Meulenaer

MLOps: challenges faced when deploying machine learning models to production

The abundance of behavioral data online and cutting-edge computational techniques seem to promise an easier way to explain and model complex human behavior. However, elegant math theories and multiplex models are largely disconnected from social science theories. There is a divide between “traditional” social scientists and data scientists in approach to data analysis.
I offer ways to address this divide in order to further advance computational social science and give examples of use of Python in social science research, from data collection to data analysis.

Dr. Inga Popovaitė

How I Learned to Stop Worrying and Love Python

Lightning Talks 2

This talk introduces a new workflow for building your machine learning models using the capabilities of modern databases that support machine learning use cases natively. There is an overview of how machine learning models are being created today to how they could look in the near future by utilising the features provided by current databases.

Developing machine learning models involves the use of data to identify patterns that would help solve business problems. Over the years as the scale of data increased, data started to get stored in databases. The model building workflows would typically fetch the data from the databases, perform some transformations to create features and use them to train the models. In some cases, these features would get stored in databases known as feature stores for reuse. To infer the model output in real-time, typically, there would be a small service or an API endpoint that would be deployed to get the results to the consumers.

As these use cases became more common, the modern databases started incorporating features that aid in building machine learning models. This talk covers some of the features provided by some of the databases including common models like linear regression, image classification, text processing, support for functions with custom models, etc. Apart from these features, many of them also make it easy to deploy the model without needing an external service for the inference. Instead, they provide native interfaces for inference like querying in SQL like languages.

This talk includes an example of how to build your custom model in Python and then include it inside your Couchbase database making inference a matter of using database queries. The example would help to understand some of the capabilities of modern databases in building machine learning models.

Nithish Raghunandanan

Select ML from Databases

Okay, so that's plainly preposterous. How is int related to list at all, much less float to tuple? Sure, int is a collection type—just one that you can't iterate over—wait, what? int is a collection type‽ What's going on here?
Join us for this keynote to find out what this all means, and why it all matters.

It's easy to get caught up in all the superficial details we are presented with or forced to learn in order to get our work done. We may even relish the opportunity to collect hundreds of little facts that we can show off to our colleagues and coworkers.

But by the fifth time you show someone the four or five or six ways to merge two dictionaries, or the eighteenth time you tweet about why hash(-2) == -2 but also hash(-1) == -2, or hour six in your explanation of why it makes sense that int is actually a collection type, well, by that point you really start to wonder: why does this all matter? Does it all matter? How does this provide any real value to me in my life?†

And yet, lurking behind each one of these seemingly superficial questions is a great depth of knowledge and understanding, spanning topics that are, in fact, immediately relevant to you, to your work, and to your life. You just have to find them.

In this keynote, we'll discuss a deliberate process and a structured framework for distinguishing “knowledge” and “meaning” from “mechanics” and “detail,” and see how each one of these nonsensical or pointless questions can be the starting point for a richer understanding of the code we write and the world of programming we inhabit.

† It took less than fifteen minutes for Thomas Caswell (Matplotlib, Bluesky) to reach this point when I started going on about such things, and he is to be given credit for this bizarre yet surprisingly apropos analogy. `list` is in fact similar to `int` in the way that `float` is to `tuple`.

James Powell

`int` is to `list` as `float` is to `tuple`

Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here. Nothing to see here.

Aidis Stukas

Introduction

School of Robotics (VšĮ Robotikos mokykla) has been looking for ways to teach a.i. for the younger generation for many years. After receiving Erasmus+ grant and gathering an international team - we have made 3 distinct courses mostly using Python. One - to control robotic arm for pick and place tasks on a conveyor belt. Another to make computer game characters using Unity ML Agents. And the third to use NVIDIA Jetbot robotic cars with cameras for self driving experiments.

A.i. based tools are gaining popularity and acceptance. Many are starting to request them in most work situations. This made a.i. programmers most sought after. At the same time a.i. development tools and hardware have become more forgiving and easier to implement. This allowed our organization to write courses of making actual computer vision applications with real world scenarios that are approachable for vocational school students starting from 9th grade.

In this talk we will share our impressions with making the courses and testing them. We will share the structure of the parts that we have written using Unity ML Agents and using Python in Unity.

We will also share the access to the courses free for educational use at gaminu.eu

Paulius Briedis

Darius Grigaliūnas

How we taught a.i. to schools in Finland, Malta, Kaunas

In this talk, Ugnė Ubartaitė will take you through her journey in the mythical programming world of 'Angis'. She will review how different points were presented by the creators. From storyline and user interface to complexity of the problem-solving and how different Python programming concepts were presented for teenagers.