MOPS - Meetup #7

Hi Pugaholics!
We are pleased to invite you to the sixth edition of the MOPS meetup which is coming back to Allegro’s office in Warsaw.
The formula is similar to previous editions, about 30-minute practical talks followed by question-and-answer sessions with networking afterward.
The entire event and all talks will be held in English.

Key details:
Location: Allegro office, 8th floor of Fabryka Norblina in Warsaw (entrance Plater 3, from Żelazna Street) 🏢
Insightful talks 💬
Knowledge exchange during networking with pizza 🍕

Plan of the meeting:
18:00-20:00 - Presentations:

1. Distributed deployment of LLMs on a Datacenter Scale (Piotr Tarasiewicz)
2. Why ReAct Fails for Smaller Language Models (and How to Fix It) (Bartosz Szaniecki)
3. Lego, Latency, and Life: Building MLOps on a Budget (Agnieszka Rybak)

20:00-22:00 - Pizza + Networking

Presentations:

• Distributed deployment of LLMs on a Datacenter Scale - This presentation introduces a framework for efficiently serving large language models in a distributed environment. The framework covers key components such as optimized data transfer, engine improvements, and a Rust-based distributed runtime. It also includes smart routing and planning mechanisms to manage workload distribution and resource allocation effectively. The framework is designed to be deployable on Kubernetes, providing a practical solution for scalable LLM inference in production. We will discuss the technical details and how this framework can be used to improve the performance and manageability of LLM deployments.
• Why ReAct Fails for Smaller Language Models (and How to Fix It) - Traditional ReAct pipelines struggle with smaller models, leading to slow reasoning, high token costs, and inefficient execution. This session begins by analyzing common issues in standard ReAct workflows, demonstrating how their complexity and rigid step-by-step reasoning create bottlenecks for ≤8B parameters models. A streaming-first ReAct approach is introduced, enabling real-time action execution and structured retrieval for better efficiency. We compare standard vs. optimized ReAct, explore LlamaIndex enhancements, and discuss Pydantic AI architectures. Attendees will learn to build faster, modular RAG agents without unnecessary complexity.
• Lego, Latency, and Life: Building MLOps on a Budget - Most of the guides on MLOps focus on companies with loads of money and applications with a lot of traffic. When deploying a small model with no funding, your constraints might be a little different. The costs are growing, the maintenance is getting in a way of your personal life and the clients complain about the latency. Everything is a trade-off. During this presentation we’ll walk through the path of building MLOps solutions for a website classifying lego bricks, what worked and what didn’t, what was a good optimization and what was the overkill.

Speakers:

• Piotr Tarasiewicz - a senior software engineer in AI at NVIDIA, specializing in Generative AI inference performance and disaggregated serving solutions. He holds a Master's in Machine Learning from University College London and a Bachelor's degree in Robotics from Warsaw University of Technology.
• Bartosz Szaniecki - a software developer at CloudFerro S.A. specializing in AI solutions, big data analysis, and EO data retrieval. His expertise spans LLM optimization, data metrics, and scalable computing, bridging AI and big data to deliver efficient solutions for cloud.
• Agnieszka Rybak - 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 - yes, the one you’re about to join.

Find us on Linkedin
If you’re interested in presenting, just write to us on LinkedIn!
Hope to see you there!