Lynn Langit

Reflections on Graph Foundation Models -https://lnkd.in/g_edj269 Understanding the semantic relationships in enterprise data remains a persistent challenge—especially when complex domains are reduced to fragmented, table-level representations in traditional relational databases. Traditionally the data model would have emerged post breaking down a domain into its constituent entities and relationships, encoded within a relational schema and organized into segregated domains across the enterprise on top of the core business entities. Ideally, the ER model reflects the domain. However, the physical implementation in relational databases often reduces interactions to isolated table queries or joins tailored to specific use cases—aggregations, metrics, groupings, etc. This tends to obscure (or sometimes erase) the deeper semantic connections embedded in the data model— specially if less thought is given to how data was originally generated, and more emphasis is placed on which source tables to extract from - irrespective of the underlying data generation process. From a dynamic system modelling viewpoint, while the ER schema represents the system's behavioral response to real-world scenarios (i.e., a data-generating process), downstream development may often treat source tables as static silos, detached from the context and logic that produced them. Moreover from an ML standpoint, building datasets off one or two tables overlooks these richer domain semantics the data model expresses. I think this is where Graph Foundation Models (GFMs) offer promise. By abstracting the enterprise data model into a graph, GFMs can potentially reason over relational structures more naturally compared to present-day LLMs (which excel at free-form data). For example, a GFM-based link prediction operator could potentially replace complex joins across tables, or even support schema evolution and legacy schema redesign (a major source of effort in legacy modernisation) in the future. A bold vision could be to rethink the ML pipeline itself — i.e instead of extracting data from a few tables and training a model per use case, a GFM trained on the enterprise database (or as context) could become a general-purpose model for classification, prediction, synthesis, and reasoning. In essence, a foundation model for the enterprise relational database. #GraphFoundationModels #GraphNeuralNetworks #RDBMS #DataArchitecture #MLPipeline #LegacyModernisation #EnterpriseArchitecture #ArtificialIntelligence

10

1 Comment

Foundry isn’t just another cloud endpoint. It’s where tens of thousands of enterprises already build, deploy, and govern AI agents at scale — with 11,000+ models, built-in security, and connectors to 1,400+ enterprise data sources like SharePoint and Fabric. Getting the newest Opus (4.6) inside that ecosystem means you can plug the most capable coding model into enterprise-grade guardrails, benchmark it against alternatives in the same catalog, and ship production agents …. without stitching together a dozen tools.

5

Google is releasing fully-managed Model Context Protocol (MCP) support for Google and Google Cloud services. We are launching with support for key services to ground AI agents in the real world and reason over enterprise data: - Google Maps: Connects agents to trusted geospatial data. - BigQuery: Enables agents to interpret schemas and query enterprise data securely. - Google Compute Engine (GCE) & Google Kubernetes Engine (GKE): Empowers agents to autonomously manage infrastructure and container operations. This is a critical step in making agentic AI easier to build and more powerful for everyone. Read the full announcement to learn more about our commitment to building the best ecosystem for AI to thrive. #GoogleCloud #AI #AgenticAI #MCP #Developer #MachineLearning #API

41

On-Demand Webinar: Accelerating ISV Innovation with Cortex Code ISV partners are increasingly looking to bring AI into their development workflows—but most coding tools weren’t designed for teams building with data. Cortex Code is Snowflake’s native AI coding agent, built to help data teams develop faster while keeping data secure inside the Snowflake platform. What You’ll Learn: ❄️ What Cortex Code is and why it matters for ISV partners building on Snowflake ❄️ Three common adoption patterns we’re seeing with ISV partners: Build Faster, Empower Customers, Prove Value with demos and POCs ❄️ How to encode domain expertise into reusable AI workflows using Skills and MCP ❄️ A live demo from Fivetran showing Cortex Code in action Check it out: https://lnkd.in/gNa9QndP Sam Mittal Randy Pettus Kelly Kohlleffel

13

🚀 GPT-5.2 Codex is now available in Microsoft Azure AI Foundry Microsoft just announced GPT-5.2-Codex, bringing enterprise-grade code generation and reasoning capabilities directly into Azure AI Foundry. What stands out to me is not just the model upgrade, but the enterprise posture behind it: security, compliance, governance, and integration designed for real production environments, not demos. For teams building developer tools, automations, or AI-assisted engineering workflows, this is another step toward making AI a first-class engineering capability, not an experiment on the side. Worth a look if you’re thinking seriously about secure, scalable AI in the software development lifecycle. https://lnkd.in/evicP86i

8

1 Comment

#Snowflake can now store #ApacheIceberg managed tables natively in #OneLake without any additional storage resources using the new #MicrosoftFabric Snowflake item. For data already stored else ware, Snowflake mirroring will now automatically create and managed shortcuts for your Iceberg tables! https://lnkd.in/eYJ8iKsE

111

The TMDL view makes it easy to reuse and share semantic model objects across different semantic models. We've just launched an official TMDL Gallery in the Fabric Community - a place where you can discover and share your favorite semantic model objects as TMDL scripts, making it simple for others to reuse them. 👉 https://lnkd.in/dnzTn4gQ cc 🎯Alexander Korn, Zoe Douglas, John Kerski, Christine Payton to move your scripts from https://lnkd.in/duZ57B7M #tmdl #powerbi

435

23 Comments

#GitHub just announced #Copilot Cloud Agent (formerly known as Copilot coding agent), signaling a clear evolution from inline assistance toward cloud��based AI support that spans research, planning, and coding. https://lnkd.in/dCkbnnzB

9

Agentic Data Quality Monitoring is now in public preview on Databricks, powered by Unity Catalog: https://lnkd.in/gBYMr8hi. This next generation of data quality monitoring looks at how your data is used and produced to automatically flag and prioritize the most important issues without needing lots of manual rules and thresholds. Just one example of how AI can transform all aspects of data engineering.

2,904

77 Comments

🚀 Excited to share the Azure AI Travel Agents sample application from Microsoft! This flagship open-source project showcases the next level of GenAI in travel solutions: orchestrating multiple specialized AI agents for real-time, personalized travel planning and assistance. ✨ What makes it unique? ☑️ Agentic architecture: Multiple AI agents (each with domain expertise), coordinated by LlamaIndex/TSC# ☑️ Powerful integrations: Built on Azure OpenAI, functions, communication services, Cosmos DB, and more ☑️ Hands-on sample: Ready to deploy and extend for real-world travel, hospitality, or concierge solutions ☑️ Truly cross-functional: Supports .NET, Python, TypeScript/JavaScript, and Java components for end-to-end workflows Whether you’re a cloud architect, AI engineer, or developer, this repo is a fantastic resource for learning about multi-agent orchestration, serverless container deployment, and next-gen AI use cases. 🔗 Check out the repo: https://lnkd.in/etRzejRx Shoutout to the MS team and all contributors for enabling innovation in the travel tech space! #AzureAI #OpenSource #TravelTech #AIOrchestration #MCP #CloudNative #GenAI #Developers

4

1 Comment

Prompt Orchestration Markup Language (POML) is a great addition from Microsoft that aims to bring structure and maintainability to prompt engineering for Large Language Models (LLMs). 𝐊𝐞𝐲 𝐅𝐞𝐚𝐭𝐮𝐫𝐞𝐬 𝐨𝐟 𝐏𝐎𝐌𝐋: • Structured Prompting: Similar to HTML, POML uses a tag-based structure with semantic components like <role>,<task>, and <examples> to organize prompts, which improves consistency and maintainability. • Diverse Data Types: You can embed various files, like PDFs and audio, directly into prompts. • Separation of Concerns: Its CSS-like styling system allows you to change a prompt's visual appearance without altering its core logic, paralleling the "view layer" in web development. • Templating Engine: POML includes a templating engine with support for variables, loops, and conditionals, enabling the creation of dynamic, reusable prompts. • Tooling and Integration: Microsoft offers a VS Code extension for syntax highlighting and real-time previews, along with a Python SDK. 𝐆𝐢𝐭𝐇𝐮𝐛 𝐑𝐞𝐩𝐨: https://lnkd.in/e6wECB-T 𝐏𝐎𝐌𝐋 𝐟𝐨𝐫 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐫𝐬: POML provides a familiar, code-like experience for professional developers. • Improved Developer Experience: The VS Code extension streamlines the workflow with features like syntax highlighting and auto-completion. • Version Control: As plain text files, POML files can be easily managed with version control systems like Git, facilitating team collaboration and change tracking. • Multi-Model Orchestration: POML simplifies working with multiple LLMs by allowing you to define which prompts go to specific models, reducing vendor lock-in. 𝐏𝐎𝐌𝐋 𝐟𝐨𝐫 𝐂𝐢𝐭𝐢𝐳𝐞𝐧 𝐃𝐞𝐯𝐞𝐥𝐨𝐩𝐞𝐫𝐬: POML's human-readable, markup-based syntax abstracts away complexity, empowering individuals without a formal coding background. • Low-Code/No-Code Approach: The intuitive language allows citizen developers to understand and contribute to prompt design. • Template Reusability: Pre-built POML templates enable non-developers to create sophisticated AI workflows without having to start from scratch. 𝐖𝐡𝐲 𝐏𝐎𝐌𝐋? Unstructured, "on-the-fly" prompts lead to the following issues: • Lack of Version Control: There's no history of changes or a way to revert to a previous version. • Inconsistency: The use of different versions of a prompt across teams results in inconsistent outputs. • No Governance: Auditing the exact instructions given to an LLM is impossible, posing a compliance risk. • Debugging Nightmares: It's extremely difficult to debug opaque, untraceable prompts. 𝐏𝐎𝐌𝐋 𝐚𝐬 𝐚𝐧 𝐞𝐧𝐚𝐛𝐥𝐞𝐫: POML transforms ad-hoc prompt practices into a formal, auditable process. Centralized, Version-Controlled Prompt Repository: POML files (.poml) are stored in a centralized repository (e.g., Git), immediately solving the versioning problem.

18

Jack Brzezinski and I “have observed a significant emergent change in the metadata neogenesis process when the process is informed primarily by domain knowledge rather than database payloads. This is a significant effect that shifts the question-answering process to a “research” type of processing. In this discussion, we build on the metadata structures from Vince Kellen, Ph.D. His “AI Metabolism of Knowledge” is inspired by biological systems. “ #ai #sql #texttosql #databases https://lnkd.in/g3nEXXjS

11

Learn how to use the ES|QL query builder, a new Python Elasticsearch client feature that makes it easier to construct ES|QL queries using a familiar Python syntax: https://gag.gl/SLMca8 #ElasticSearchLabs

5

Know where you are in your AI journey. Grow gradually, slowly and deliberately. Don’t leapfrog the stages. Stop blaming AI for poor quality output. Most of the time, it’s you not understanding which stage fits your use case and then not architecting it the right way to solve it.