Abstract

We're bringing LLMDay + Data to Tel Aviv for the first time ever in partnership with Microsoft and TeraSky to focus on the intersection between LLMs, AI and Big Data Engineering. Expect excellent talks on everything from LLMs to data engineering, analytics and how all of this powers what's coming next with AI. Join us at the Microsoft Reactor, in Midtown Tel Aviv with 200 data and LLM enthusiasts - and get your talks submitted, CFP is open until October 17th!

Bio

Sharone Zitzman, is a developer relations professional and an open source community builder, who likes to work with engineering teams that are building products that developers love. Having built both the DevOps Israel and Cloud Native & OSS Israel communities from the ground up, today she spends her time at her company RTFM Please, building dev-first adoption strategies to bring excellent technology and people together. You can find her at rtfmplease.dev, or on Twitter & Github as @shar1z.

Abstract

Bio

Yev Berman is Director of Sales Engineering at TeraSky, where he helps enterprises design and adopt cloud native, hybrid cloud and automation platforms. Before his current role he led TeraSky's Private Cloud Delivery group and its Hybrid Cloud & Automation team, and he was named a VMware vExpert in 2024.

Abstract

AI coding agents do well on local tasks like fixing a function or writing a test. Ask one how a large system fits together, or what breaks if you change a shared type, and it falls back to grep and reading files one at a time. On a codebase spread across many repos, that is slow, burns tokens, and misses connections the agent never opens. This talk covers a different approach: turn the codebase into a graph and let the agent query its structure directly. Functions, calls, types, HTTP endpoints, and links between repos become nodes and edges the agent can traverse in a single query. The working example is Auryn, which parses 15 languages into a FalkorDB graph, updates it on every commit, and serves it to any agent over MCP, with semantic search so the agent can find code by what it does and not only by its name. I’ll cover how source code maps to a graph, how to keep the graph current as the code changes, and where the graph helps most and where reading files is still the right call. Then a live demo: an agent answering architecture and impact questions across repos, compared with the same questions answered by reading files.

Bio

Roi Lipman is CTO and co-founder of FalkorDB, the low-latency graph database built for GenAI and GraphRAG workloads. He created RedisGraph and led it as lead architect at Redis for eight years, and brings over 20 years of database engineering experience from Forter, StreamRail, AVG and the Israel Intelligence Corps.

Abstract

Our production AI agent went through eight complete rewrites before its architecture finally held. Some fixed things we’d gotten fundamentally wrong. Others happened because the ground moved: a model or platform release turned last quarter’s hard-won infrastructure into a line in someone else’s changelog.

When the model labs keep swallowing whole layers of the stack, every team building agents faces the same dilemma: build now, knowing you may be eclipsed in six months, or wait and fall behind. “Build vs. buy” stops being a milestone and becomes a decision you revisit every quarter.

In this honest post-mortem, we’ll walk through what broke, the stack that earned its place (LangGraph for orchestration, AWS Bedrock and Google Vertex for model flexibility, Temporal for workflow reliability), and which of those choices we’d still bet on a year from now.

The most durable lesson was the least glamorous: discoverability. Making the agent useful across a complex product meant building a common query interface that makes every capability, metric and data point legible to it in a consistent, verifiable way. That layer turned hallucination into an architecture problem with an architecture solution, and it’s the one part of our system no lab can ship for us.

Attendees building agents in production will leave with a practical way to separate the layers worth owning from the ones worth renting, and what it concretely takes to make agent outputs evidence-backed rather than taken on faith. Sometimes the best engineering outcome is deleting code you were proud of. The hard part is knowing which code.

Bio

Avital Tamir is a generalist, software developer, twin-daddy and former pro musician. He taught himself BASIC at age 6 when he got his first x86 and has been tinkering with computers ever since. As a musician Avital released several studio albums and toured the world, before shifting to a full-time career in tech. Avital is the creator of the Cyphernetes query language and an avid user of everything AI.

Abstract

Every new generation of hardware is bigger and faster than ever, yet much of that capacity is lost before it does useful work. It goes to threads waking up, waiting on locks and handing work to each other.

The cost of switching shows up everywhere, from big data processing to people: switching between tasks can cost up to 40% of productive time. Whether the brain is silicon or human, switching context is expensive.

Thread-per-core and run-to-completion designs are well established in systems where latency matters most. Give each core its own work, keep data close to the thread that uses it, and let every task finish without handoffs. The results can be dramatic. But what does it take to apply these ideas to a large distributed data platform?

This talk shares Cato’s experience rebuilding our real-time pipeline around focus. We’ll cover what we had to rethink about threads, caches, queues and shared state, where the approach turned out simpler than expected, and the tradeoffs we accepted along the way.

Bio

Noa is a Data Engineering Team Lead at Cato Networks with a big passion for infrastructure. Over the past 5 years she has been working with data and helping data teams to successfully handle the growing amount of data collected from the Cato Cloud. Her work focuses on building the new Cato Data Platform, to ensure it supports the organization growing demand for data in scale and use cases with efficient and cost-effective solutions, all while keeping up with Cato’s growth and achieving outstanding performance.

Abstract

Our platform relied on Amazon Timestream to store the metrics behind customer-facing dashboards and incident investigations. When AWS announced the service’s end of life, we had to migrate 5.9 billion metric rows a day to ClickHouse without taking queries offline or compromising the integrity of the data.

This session breaks down the migration from the decisions that shaped the new schema and partitioning strategy for high-cardinality Kubernetes metrics to the validation process that gave us confidence to cut over. We ran dual-write and dual-read paths and compared 10.3 million production queries over 14 days to establish parity before switching over. We’ll cover where results diverged, why the migration took roughly eight weeks longer than estimated, and how the actual cost and performance gains compared with our expectations: approximately 85% lower costs and median queries around 20 times faster.

Attendees will leave with a reproducible method for validating metrics-store migrations, the ClickHouse settings that mattered at this scale, and a practical cutover approach for moving production observability data without disrupting customers.

Bio

Nir Adler is an Innovation Engineer on Komodor’s CTO team, driving AI initiatives like Klaudia, the AI SRE transforming Kubernetes troubleshooting. A self-taught developer whose journey began with Lego Mindstorms, Nir’s career spans cybersecurity, infrastructure design, and open-source contributions, with senior roles at Palo Alto Networks and Snappy Gifts. Blending creativity with deep technical expertise, he helps teams build secure, scalable, and impactful products  ( and always with the loyal supervision of his dog, Maze.)

Abstract

Most AI architectures depend on a simple trick: compress data, average metrics, cluster vectors, and let the LLM summarize. But in domains like construction, manufacturing, or physical operations, averaging anything lies to you. 99% progress means nothing if the missing 1% is a structural beam. This talk covers an anonymized case study of building an LLM system over non-aggregatable data—why vector search and rollups failed, and the architecture that actually made it work.

Bio

Liza Katz is the CEO of Neradot and a seasoned consultant with 20+ years of experience in large-scale software development, with a strong foundation in data engineering, observability, and search — including deep expertise in Elasticsearch and OpenSearch, specifically in the context of AI and search-driven applications.

Abstract

Did you also join the wave of building a second brain with a folder of markdown, an instructions file, and maybe Karpathy’s LLM Wiki pattern on top? I built one to run my actual job (leading a team, shipping research, event speaking) with an agent wired into my email, chat, calendar, meeting notes and work tracker. Six weeks in, it had hundreds of tidy notes but it was quietly failing. Review items nobody saw. Decisions frozen for a month. Notes written once and never touched again. Writing knowledge down turned out to be the easy part.

This talk is the field report on what came next. Why I dropped PARA for an entity graph. The workflows and scheduled jobs that push what needs me, take my replies in plain words, and close what’s done with evidence. The deterministic scripts that keep an LLM-maintained system honest. The guardrails for an assistant that reads untrusted email and drafts public posts. You’ll see what worked, where it still falls short, what it would take to scale it into a real personal AI assistant, and a live demo of the open kit you can use to build your own.

Bio

Liran Tal is an AI security researcher, open-source maintainer, and author of books on Node.js security. A GitHub Star and recipient of the OpenJS Foundation’s Pathfinder for Security award, he leads Developer Relations at Snyk and helps developers build more secure software.

Abstract

Most enterprise knowledge doesn't answer questions — it sits locked in semantic models, dashboards, and databases, waiting for someone who knows DAX or SQL to ask. This workshop shows you how to connect Claude AI to a Microsoft Fabric semantic layer using the Model Context Protocol (MCP), turning your organization's data into a conversational Company Brain that any business user can talk to. We'll build an end-to-end agentic RAG architecture on Azure: indexing your Fabric schema into Azure AI Search, generating embeddings with Azure OpenAI, and deploying a Remote MCP Server on Azure Container Apps that exposes structured tools — schema retrieval, DAX execution, and dataset discovery — to Claude. You'll see how a business user can ask "what were our top products by revenue last quarter?" in plain English and get an accurate, governed answer in seconds, with no DAX knowledge required.

We'll also cover the security layer: OAuth 2.0 with Azure Entra ID, On- Behalf-Of token flow, and automatic row-level security — so your

Company Brain respects the same governance policies as your Fabric workspace. You'll leave with a working architecture pattern you can adapt

to your own data models, and a clear understanding of when RAG-over- schema beats fine-tuning or search-only approaches.

Bio

Haggai Zohar is a Data Tech Lead at TeraSky, where he designs cloud data platforms and the AI systems built on top of them. A Technion graduate (cum laude), he holds AWS Machine Learning Specialty and AWS Solutions Architect certifications and works across AWS, Azure and Microsoft Fabric. He recently spoke on "RAG on Your Data Lake" at TLV Community's AWS DevOps 2026 conference, and he loves turning messy company data into something an LLM can actually use.

Abstract

Join our esteemed panel as we probe them (grill them actually) on what they think is coming next as the worlds of Data + AI explode - how the data engineering / science domain is impacting the AI discipline and vice versa...and what they think we all need to prepare for.

Moderated by Sharone Zitzman.

Bio

Adaya Tal is the Data Manager at ReasonLabs, where she leads a lean data team that owns the company's data infrastructure, semantic layer and the internal AI agents serving product, marketing and finance. She works at the point where data engineering meets AI adoption, turning messy data into something both people and agents can trust. Adaya co-hosts the Data Bites podcast on data and AI, runs the Data Queens community for women in data, AI, product and marketing, and mentors data professionals in the Israeli tech scene.

Einat Borohovich turns advanced data concepts into reliable systems. As a Senior Data Scientist at Rambam Health Care Campus, she brings a hands-on approach to engineering, ensuring that innovation translates into reality. Grounded in statistics, Einat's work naturally blends core data architectures with cutting-edge NLP, LLMs, and agentic AI. Always looking ahead, she is currently expanding her technical horizons into quantum computing.