Interpreting Black Boxes and Benchmarking the Machines Behind Them

🚀 Join our free Budapest Data Science Meetup!

In the first talk, Domokos Miklós Kelen from Mastercard will dive into SHAP in practice, showing how Shapley-based explanation methods are actually used in production systems, from theory and architectural choices to real-world trade-offs when explaining black-box models.

Next, Gergely Daróczi, PhD will present insights from benchmarking 2,500+ cloud server types across six providers, sharing performance results from hundreds of DS/ML/AI workloads — including LLM serving — and revealing what really matters when choosing infrastructure for scalable, efficient inference.

📅 Date & venue:

• 29 January 2026, Thursday, from 6 pm
• Mastercard Office 1132 Budapest, Váci út 26. · Budapest

Please note that the event's official language is English!

Schedule

17:45 gates open
17:45-18:15 Warmup and chit-chat
18:15-19:30 Talks
19:30 - 21:00 Drinks, snacks, networking

Explaining Production Models: SHAP in Practice

Abstract:
Since its publication, SHAP and other Shapley value-based explanation methods have become an industry standard for gaining insight into black-box models. However, there are many variants of SHAP, each with its own trade-offs. In this talk, we’ll discuss how explanation methods are used in our production system, covering the theory behind them, the system architecture, and practical considerations.

Domokos Kelen, Lead Data Scientist at Ekata, a Mastercard company. Having completed his PhD in Machine Learning recently, he joined Mastercard in 2025 to work on building, improving, and explaining fraud detection models.

----------

Benchmarking 2000+ Cloud Servers for DS/ML/AI Workloads

Abstract:
Spare Cores is a Python-based open-source ecosystem collecting, generating, and standardizing comprehensive data on cloud server pricing and performance. We have evaluated over 2,500 server types across six cloud vendors to measure performance across ~500 workloads, including suitability for serving Large Language Models from 135M to 70B parameters. We tested how efficiently models can be loaded into the memory or VRAM, then measured inference speed across varying token lengths for prompt processing and text generation. The published data can help you find the optimal instance type for your LLM serving needs, and we will also share our experiences and challenges with the data collection, and lastly some insights into general performance patterns.

Gergely Daroczi, PhD, is a passionate R/Python user and package developer for two decades. With over 15 years in the industry, he has expertise in data science, engineering, cloud infrastructure, and data operations across SaaS, fintech, adtech, and healthtech startups in California and Hungary, focusing on building scalable data platforms. Gergely maintains a dozen open-source R and Python projects and organizes a tech meetup with 1,800 members in Hungary – along with other open-source and data conferences.

Sponsors and partners

• Ekata, a Mastercard company
• Budapest Data Science Meetup