Gennadi Lembersky, PhD
Principle Research & Data Science Manager
NICE
NICE’s LLM Journey: Challenges and Insights
________________________________________________________________________
A year has passed since GenAI gained widespread recognition and industry adoption. In this presentation, I will share NICE’s transformative journey with large language models (LLMs) during this period. I’ll explore emerging use-cases, address challenges, provide initial insights, and discuss my perspective on an LLM-powered future in customer experience.
https://www.linkedin.com/in/gennadilembersky/
About Gennadi Lembersky, PhD
Follow Gennadi Lembersky, PhD
Principle Research & Data Science Manager
NICE
Itamar Friedman
Co-Founder & CEO
CodiumAI
Itamar from CodiumAI showcases the latest advancements and features that enhance code generation when coupled with code testing and review. CodiumAI recently released AlphaCodium, which is a first-of-its-kind open-source tool that, with a click of a button, generates better results than the majority of professional developers in code contests.
https://www.codium.ai
About Itamar Friedman
Follow Itamar Friedman
Co-Founder & CEO
CodiumAI
Nir Makmal
Chief Architect | M.Sc, CS | AI/ML/GenAI
Amdocs
Revolutionizing Telco GenAI Agents with Amdocs amAIz: An E2E Journey of Creating GenAI Agents
______________________________________________________________________
We will share one of our interesting real-world GenAI Agents Telco Enterprise-grade use cases, where we navigated through complex challenges using innovative data and AI techniques. Discover how we significantly enhanced model accuracy and reduced token consumption, paving the way for efficient GenAI Agents with the Amdocs amAIz GenAI platform.
About Nir Makmal
Follow Nir Makmal
Chief Architect | M.Sc, CS | AI/ML/GenAI
Amdocs
Or Dagan
VP Product | Foundation Models
AI21 Labs
AI21 Jamba: going beyond Transformers
____________________________________________________________________
Abstract:
In March, AI21 released Jamba, the world's first production-grade Mamba-based LLM. Jamba is based on a novel architecture that merges both SSM and Transformers, benefitting from the best of both worlds. Or Dagan, VP Product Foundation Models at AI21, will explain the architectural differences, the potential gains and the lessons learned when building this model.
https://www.ai21.com/jamba
About Or Dagan
Follow Or Dagan
VP Product | Foundation Models
AI21 Labs
Naama Damty
VP, Chief Architect, CX Division
NICE
About Naama Damty
Follow Naama Damty
VP, Chief Architect, CX Division
NICE
Ofir Yakobi, PhD
Data Science Group Lead
NICE Actimize
Leveraging Word Embedded Vectors and Similarity Scores to Combat Financial Crime: Opportunities and Challenges
___________________________________________________________________
Effective data representation has always been crucial in machine learning applications, from traditional statistical models to modern deep learning frameworks.
The representation of tabular financial data as narratives, and subsequently as embedded vectors, has opened new ways for enhancing accuracy and insights in financial crime detection.
Embedded vectors enable the handling of data-related challenges like high dimensionality and sparsity, and in some cases, they allow for the capturing of nuanced patterns and relationships that might otherwise be missed.
I will explore the use of word embedded vectors and similarity scores to identify and prevent financial crimes.
About Ofir Yakobi, PhD
Follow Ofir Yakobi, PhD
Data Science Group Lead
NICE Actimize
Elad Degany
Chief Data Officer | Driving Data & AI Transformations
Elal
AI Transformation in Elal
About Elad Degany
Follow Elad Degany
Chief Data Officer | Driving Data & AI Transformations
Elal
Yoav Avneon, PhD
Data Science Group Lead
NICE Actimize
Opportunities to Embed GenAI in X-Sight for Advancing FinCrime Solutions: A Dual Use-Case Exploration
About Yoav Avneon, PhD
Follow Yoav Avneon, PhD
Data Science Group Lead
NICE Actimize
https://www.iahlt.org
About Avner Algom
Follow Avner Algom
GM
IAHLT
Yael Mathov, PhD
Staff AI/ML Researcher
Intuit
Staff AI Security Researcher at Intuit
_____________________________________________________________________
As the adoption of Large Language Models (LLMs) surges within the industry owing to their exceptional abilities in deep comprehension of complex text, their unique capabilities come with unique risks and challenges. With their ability to mimic human-like reactions, LLMs could have potential implications for the security and reliability of the systems they power.
In this talk, we will investigate potential vulnerabilities within LLMs that could be exploited, illustrating how these could be used to manipulate models in operation. We will then share a set of strategies and proactive measures that data scientists at Intuit implement to bolster our AI-powered applications against malevolent attacks. This approach helps multiple teams harness the power of LLMs while mitigating their inherent security risks.
About Yael Mathov, PhD
Follow Yael Mathov, PhD
Staff AI/ML Researcher
Intuit
Kfir Bar, PhD
Senior Lecturer at Reichman University and Chief Scientist at Babel Street
Reichman University
About Kfir Bar, PhD
Follow Kfir Bar, PhD
Senior Lecturer at Reichman University and Chief Scientist at Babel Street
Reichman University
Gilad Fuchs, PhD
Senior Applied Researcher
eBay
Embeddings-Based Retrieval for Effective Pricing in E-commerce
_____________________________________________________________
Helping sellers price their listings is an important and challenging task for E-commerce marketplaces. To help the seller gain trust in the recommended price, a collection of supporting similar listings are retrieved and provided along with their prices. We address the problem of retrieval-based price recommendation using a novel approach, which enables a trade-off adjustment between semantic similarity and price accuracy. Balancing the two required since, based on our study, retrieval of semantically similar listings does not guarantee pricing accuracy. In contrast, a price-accuracy driven approach may produce less semantically supporting listings. We also suggest a third method - training a Multi-Task network which learns in parallel both semantic similarity and a pricing-based objective. Framing the solution as a Multi-Task network unfolds the ability to control the balance between explainability and accuracy, thus providing a powerful tool to precisely tailor the correct pricing solution to different real world business use cases.
About Gilad Fuchs, PhD
Follow Gilad Fuchs, PhD
Senior Applied Researcher
eBay
Peter Lifshits
Lead Software Architect
NICE
Text2SQL in practice
_____________________________________________________________________
Text to code in general, and text to SQL in particular is a well-known problem, with quite a lot of research, benchmarks and ratings. However, when it comes to implementing Text2SQL as a production-ready feature in a specific data environment for a specific application, there are a number of factors to consider which are usually not evaluated by researchers and not included in their benchmarks – neither in terms of quality nor performance. This presentation highlights a few practical challenges that arise from a real-world application of Text2SQL and their proposed solutions.
About Peter Lifshits
Follow Peter Lifshits
Lead Software Architect
NICE
Roi Lipman
CTO & Co-founder
FalkorDB
Knowledge Graphs VS Vector DBs in RAG solutions
______________________________________________________________________
AI solutions relay on context retrieval to query private data and combat hallucinations, the big question is which data store is better suited for context extraction? in this talk we'll compare the two alternatives and examine their end results.
About Roi Lipman
Follow Roi Lipman
CTO & Co-founder
FalkorDB
Guy Eyal, PhD
Senior Manager NLP
Gong
About Guy Eyal, PhD
Follow Guy Eyal, PhD
Senior Manager NLP
Gong
Lee Twito
GenAI Lead
Lemonade
RAG pain points and solutions building customer-support GenAI agent
______________________________________________________________________
Lee presents practical approaches for improving Retrieval-Augmented Generation (RAG) in production-grade LLM applications. Lee will discuss how to overcome common challenges in building a GenAI agent that autonomously handles customer support tickets, while ensuring broad customer intents coverage and precise responses. The talk will include insights on both the query and index parts of the Llm app pipeline.
About Lee Twito
Follow Lee Twito
GenAI Lead
Lemonade
Elik Sror
Algorithm Team Lead (NLP-GenAI)
WSC Sports
Summarizing sport games using NLG: Real use of LLMs in production
______________________________________________________________________
The landscape of digital sports content has undergone significant transformations in recent years, accompanied by a substantial shift in the consumption patterns of sports enthusiasts. This content now encompasses game highlights, recaps, press conferences, and more, all presented in descriptive summary formats. In this presentation, we will explore our methodology for autonomously generating game recaps through the application of Large Language Models (LLMs) and prompt engineering techniques. The discussion will encompass the diverse challenges encountered throughout this process and effective strategies for overcoming them.
About Elik Sror
Follow Elik Sror
Algorithm Team Lead (NLP-GenAI)
WSC Sports
Aviv Slobodkin
PhD student in CS @BIU, specializing in NLP
Bar-Ilan University
Attribute First, then Generate: Locally-attributable Grounded Text Generation
_____________________________________________________________________
Recent efforts to address hallucinations in Large Language Models (LLMs) have focused on attributed text generation, which supplements generated texts with citations of supporting sources for post-generation fact-checking and corrections. Yet, these citations often point to entire documents or paragraphs, burdening users with extensive verification work.
In this line of work, we introduce a locally-attributable text generation approach, prioritizing concise attributions.
Our method, named "Attribute First, then Generate", breaks down the conventional end-to-end generation process into three intuitive steps: content selection, sentence planning, and sequential sentence generation.
By initially identifying relevant source segments ("select first") and then conditioning the generation process on them ("then generate"), we ensure these segments also act as the output's fine-grained attributions ("select" becomes "attribute").
Tested on Multi-document Summarization and Long-form Question-answering, our method not only yields more concise citations than the baselines but also maintains - and in some cases enhances - both generation quality and attribution accuracy. Furthermore, it significantly reduces the time required for fact verification by human assessors.
About Aviv Slobodkin
Follow Aviv Slobodkin
PhD student in CS @BIU, specializing in NLP
Bar-Ilan University