Machine Learning Prague 2026

Large language models are revolutionizing how we build intelligent systems, but deploying them on massive datasets in production brings serious challenges — from runtime bottlenecks and spiraling costs to maintaining output quality at scale. 
In this talk, I’ll walk through a real-world case study of running LLMs over big data pipelines, sharing lessons on optimizing inference performance, managing compute costs, and ensuring accuracy.

Every team deploying LLMs eventually runs into the same question: “What’s the optimal setup for my model?” But as soon as you start moving beyond toy examples and into real production traffic, you discover an uncomfortable truth: optimal doesn’t exist. It shifts. It depends. And every decision — GPU type, model variant, framework, batch size, prompt shape, concurrency — creates a new performance reality.

In this talk, we expose the hidden dynamics behind LLM performance through the lens of 300,000+ large-scale benchmarks executed across GPUs, frameworks, models, and scenarios. These aren’t synthetic micro-tests. They’re high-load, stress-tested, real-world workloads run with an automated benchmarking pipeline capable of executing hundreds of thousands of experiments with almost no operational cost.

The rapid adoption of AI systems, from web applications and decision-making systems to autonomous agents and LLM-based solutions, is leading to a significant shift in the cybersecurity risk profile. In practice, many organizations deploy machine learning models as just another application component, overlooking the fact that AI introduces an entirely new class of vulnerabilities, attack vectors, and systemic risks that cannot be adequately addressed using traditional cybersecurity approaches alone.

In this presentation, I analyze the security risks associated with AI implementation by examining both its connection with classical application vulnerabilities and the emergence of novel threat categories. Drawing, among others, on the OWASP Top 10 for Large Language Model Applications and the OWASP Machine Learning Security Top 10, I show that the most significant security threats stem from erroneous architectural, operational, and human assumptions, particularly those related to trust, system boundaries, and the behavior of learning systems.

The evolution from simple RAG workflows to production-ready agentic AI requires solving three fundamental challenges: providing agents with precise context, maintaining observability at scale, and adapting rapidly to new models. This talk shares our journey building a comprehensive AI platform that addresses these challenges through context ownership, systematic monitoring, and flexible model integration.
Our solution centers on three pillars: First, we own our entire search stack and integrate diverse data sources through Model Context Protocol, ensuring agents receive high-quality context precisely when needed. Second, we maintain production reliability through comprehensive evaluations, continuous monitoring, and automated red-teaming for security. Finally, our internal LLM leaderboard enables rapid evaluation and switching between the latest models, seamlessly combining commercial APIs with custom fine-tuned and self-hosted solutions to optimize performance for specific agent workflows.

Everyone building agentic systems faces the same question: how do you know your agent is doing what you think it's doing? The honest answer is deceptively simple, you look. But looking well turns out to be surprisingly hard. Agent evaluation borrows from software testing, ML benchmarking, and empirical research methods, but fits none of them cleanly. We're trying to evaluate inherently stochastic systems with tools designed for deterministic ones.

At Aisle, we build AI agents that detect and fix security vulnerabilities across enterprise codebases. This talk shares practical lessons from evaluating multi-agent systems in a complex, high-stakes domain. We'll cover the tools we use, the mental models that help, and the traps we've learned to avoid.

Tool-using agents are powerful, but most MCP integrations are still inefficient: they paste tool schemas into prompts, shuttle large intermediate outputs through context, and lose reliability as workflows grow. This wastes tokens, adds latency, and increases failure rates in multi-step tool chains.
Cloudflare and Anthropic recently proposed MCP “code mode”: instead of prompt-time JSON tool calls, the model generates small programs that call MCP tools via an API and run in a sandbox. This dramatically reduces overhead (Anthropic reports a reduction from 150,000 to 2,000 tokens in a representative workflow) while improving accuracy in chained tool use.
In this talk, we’ll present a terminal-first workflow using `mcpc`, a universal MCP CLI client, and live-demo parallel calls to multiple MCP servers, local filtering/transformation (Unix style), and “export to script” so agent interactions become reusable, auditable pipelines. You’ll leave with practical design patterns and an evaluation checklist (cost, latency, error rate) to build scalable MCP systems, independent of any single coding agent.

Identifying enzymes for specific chemical transformations remains a key challenge at the intersection of machine learning and biochemistry. We present a modular framework that integrates over 20 open-source tools to support ML-driven enzyme discovery from large sequence databases. As part of this, we developed a machine learning model to map between chemical reaction representations and the protein sequence space, enabling large-scale retrieval of candidate enzymes for a target transformation. We experimentally verify the predictions and identify multiple enzymes capable of degrading synthetic pollutants that are dissimilar to the training set, but enzyme discovery is far from solved. I’ll conclude by highlighting key challenges and opportunities, such as improved molecular encodings and model architectures for biochemical data.

Plants are an incredibly rich source of pharmaceutically relevant specialized metabolites (small molecules). However, deciphering the biochemistry that occurs in non-model plant species is a highly challenging task. My lab is developing generalized computational workflows for connecting biosynthetic gene sequences (obtained using RNA sequencing) to their downstream metabolites (detected using mass spectrometry). To this end, we have developed EnzymeExplorer, a targeted machine learning pipeline for predicting the enzymatic functions of terpene synthases directly from their amino acid sequences and DreaMS, a self-supervised foundation machine learning model for tandem mass spectrometry, which outperforms state-of-the-art methods in a range of different prediction tasks. We are further building on these foundations towards the final goal: a full computational characterization of the chemodiversity and biosynthetic potential of each plant species using easy-to-obtain experimental datasets.

“Why Good Models Fail After Deployment” explores a paradox familiar to many ML practitioners: models that perform exceptionally well during training and offline evaluation often degrade—or even cause harm—once deployed in real-world systems. This talk examines structural reasons behind this phenomenon, focusing not on model architectures, but on the environments in which models operate.

Drawing from practical experience optimizing large-scale recommendation systems at Meta (Instagram Reels) and large AI frameworks at Intel, the session highlights common failure modes such as distribution shift, feedback loops, metric misalignment, delayed effects, and biased data collection. Attendees will learn why offline metrics often fail to predict real-world performance, how optimization objectives can lead to unintended long-term system behaviors, and design principles to improve system robustness.

The talk emphasizes system-level thinking, mental models, and practical strategies, applicable across ranking, search, forecasting, and decision-making systems.

Sequential recommendation models based on self-attention, such as SASRec, are a natural fit for e‑commerce search: they can condition on a user’s full interaction history and adapt as intent evolves. At Bloomreach, we use a SASRec‑style model to re‑rank the top 200 results from an existing learning‑to‑rank (LTR) model for both keyword and category search, rather than replacing the ranking stack outright. This talk shares what we learned while turning that idea into a production system across a wide variety of customers.

Building a proprietary transformer model with tens of millions of parameters is an engineering challenge. Training it on 100TB of proprietary Counter-Strike 2 data is a logistical nightmare.

At Rankacy, we moved from a comfortable on-premise R&D setup to a full-scale AWS production pipeline to analyze complex player behaviors. It wasn't smooth. This talk skips the "AI is magic" hype and focuses on the hard reality of building and deploying custom models at this scale.

AI-driven digital avatars have great potential in the online entertainment industry, from gaming and live streaming to interactive virtual events. The key challenge has always been balancing speed and quality: the most realistic avatars are often too slow, while real-time solutions sacrifice visual fidelity. In this talk, I will share how new approaches, such as separating motion from appearance, are making it possible to create hyper-realistic, fully controllable avatars that run in real time. These breakthroughs open the door to more immersive, engaging, and personalized digital experiences, bringing online entertainment closer to truly lifelike interaction.

Face recognition is a natural human skill refined over a lifetime — yet modern algorithms now surpass us even under challenging conditions. These challenges include appearance changes caused by aging, make-up, or lifestyle; variations in camera quality, viewing angles, and environmental conditions; long-range turbulence; and, importantly, demographic differences. We will also touch on scenarios where faces are captured from longer distances up to one kilometer, where factors such as optical limitations and atmospheric effects introduce unique constraints on recognition.

In this talk, we will examine how these factors affect system performance, how to measure them, and most importantly, how to mitigate them. Since most state-of-the-art systems rely on neural networks, data quality becomes the primary driver of robustness. We will demonstrate how Innovatrics — a company with more than 15 years of experience in face recognition — approaches data: from dataset cleaning and identity clustering to domain adaptation.

3D reconstruction is a powerful and useful tool for holographic communication and an increasingly interesting topic in the field. Reconstructing the human head under everyday conditions presents a set of challenges, the most important being that the head is not a rigid object.  Neural networks have enabled significant progress in 3D reconstruction, first after 2017 with convolutional neural networks, and later, after 2020, with neural radiance fields (NeRF). Since then, NeRF-based approaches have evolved rapidly, through Instant-NGP, then Gaussian splatting, then triangle splatting, and other techniques sharing roots with these methods.  In this talk, we demonstrate the application of these techniques to human head reconstruction for holograms, discuss the specifics and evolution of particular methods, and explore the motivation for using neural 3D reconstruction in holographic communication systems.

This talk introduces Contrastive Forecasting, an unsupervised method for time series prediction that learns entirely in latent space. The approach is grounded in contrastive divergence and uses a joint-embedding predictive architecture (JEPA) to align predicted future states with actual outcomes while distinguishing them from carefully selected negative samples.

We will detail the model architecture, which combines RWKV or Transformer-based forecasters with Residual encoders. The training objective encourages accurate representation learning by pulling forecasted embeddings toward future targets and pushing them away from dissimilar contexts.

The talk will cover practical aspects of training, the design of contrastive losses for temporal data, and the handling of multivariate and long-horizon forecasting challenges. This session is intended for researchers and practitioners interested in self-supervised learning, representation learning, and time series modeling.

Retrieval is one of the main components of modern RAG applications. Many retrieval systems are based on embedding models which are used for underlying vector representations. Building a State-of-the-Art (SOTA) embedding model requires rigorous engineering, precise data curation, and architectural innovation.

In this talk, we present a technical deep dive into the creation of nvidia/llama-embed-nemotron-8b, the model that achieved the #1 rank on the Multilingual Massive Text Embedding Benchmark (MTEB). Specifically, we will reveal the architectural modifications, training objectives, and large-scale data curation strategies required for SOTA performance.

Furthermore, we will explore the transition from text-only to multimodal retrieval. Attendees will get an under-the-hood look at nvidia/llama-nemoretriever-colembed-3b-v1, a SOTA VLM-based embedding model, and learn how to adapt retrieval systems to handle complex visual data effectively.

Traditional full-text search has served us well for decades, but it carries fundamental limitations that become increasingly apparent as user expectations grow. Typos derail queries. Synonyms go unrecognized. A search for "car" fails to surface results about "automobiles." Cross-lingual retrieval remains a distant dream. And contextual understanding—grasping that a query about "products by manufacturer X" should return relevant items even when the manufacturer name doesn't appear explicitly—simply isn't possible with keyword matching alone.
 
This presentation explores a practical solution: hybrid search that combines the precision of classic BM25 full-text algorithms with the semantic power of vector embeddings. We'll examine how embeddings map text into multi-dimensional spaces where meaning, not just characters, determines similarity. This enables retrieval based on concepts rather than exact strings, opens the door to cross-lingual search, and allows the system to understand context and relationships between entities.