WiMLDS - Optimizing Latency, Observability & Image Watermarking in AI world

We are coming back with another WiMLDS meetup <3

We will meet on October 14th (Tuesday) at 6PM at Fandom Office (Baraniaka, Poznań).

We will hear three great talks in MLOps and Data Science/Research areas.
Below you can find information on our speakers:

Agnieszka Rybak
Agnieszka is a Software Engineer with many years of experience in fields such as MLOps, Software Engineering, DevOps, and Machine Learning. Besides being an MLOps Engineer at NVIDIA, she’s currently working on a side project web application Brickognize.
Previously worked on the MLOps infrastructure at Allegro and before that gained experience in companies such as Semantive, Mozilla, and Facebook. Co-founder of MOPS Community - the first in Poland MLOps meetup.

Title:
Bricks and Tricks: Optimizing Latency Across Environments

Abstract:
Imagine building an app that recognizes any LEGO brick in a photo and making sure users don’t have to wait for the result. Achieving fast predictions isn’t as straightforward as it seems - the right approach depends a lot on where the model is running and how it’s being used. Whether it’s a LEGO sorting machine or a web server handling lots of requests, whether you have money for a GPU or only for a CPU server, the obstacles and solutions can look very different. In this talk, I’ll share real-world examples of reducing latency in different setups, highlighting what worked, what failed, and why there’s rarely a single “magic fix.”

Zuzanna Gawrysiak
Zuzanna is an AI Engineer at Vestigit, where she develops deep learning methods for video watermarking to help fight digital piracy. Her main focus is computer vision, and she’s recently started a PhD at Poznań University of Technology, exploring neurosymbolic architectures for computer vision tasks.

Title:
Invisible Yet Invincible: Deep Learning Approach to Image Watermarking

Abstract:
In a world where digital content is copied and shared in an instant, how can creators protect their work? This presentation explores a new approach using deep learning to create invisible watermarks. We've built a smart AI that hides information directly inside an image's pixels without changing how it looks. The secret is that this hidden watermark is also incredibly tough—it's designed to survive the journey across the internet, including being compressed, cropped, or blurred. Our focus was on building a tool that is not only strong but also fast and practical for real-world use. It offers a powerful way for artists, photographers, and companies to prove ownership and fight back against digital piracy.

Julia Będziechowska
Senior Machine Learning Engineer at Fandom. My day to day work focuses on bridging the gap between AI experimentation and production environments, from infrastructure (AWS) through implementation to maintenance. I find myself a product-minded engineer who prioritizes finding use cases that enable organisations to leverage the power of AI, with all that it takes: building trust in AI, proving solution reliability, and establishing common understanding. After working hours, I dedicate myself to another passion of mine which is playing bass guitar and singing.

Title:
ML Observability - from Idea to Production

Abstract:
The story of turning the idea of ML Observability into a deployed solution, which took the maturity of our AI solutions at Fandom to the next level.
The established solution provides automatic detection of malformed datasets, exposes insights on ever-changing production data and democratises data audit in our organization.
I will share both key takeaways on approaching the project focused on Quality Assurance in the AI field, as well as managing its scope and requirements for the ultimate win.
Technical spoilers → EvidentlyAI, Apache Superset, Airflow, AWS

Hope to see you there,
WiMLDS Poznan team