Julie Strauss

Most AI agent samples teach you to call a model. They rarely show what happens when five agents need to coordinate, tools need to ship independently, and the whole stack needs to deploy to the cloud. Justin Yoo's Interview Coach sample fills that gap. It puts Microsoft Agent Framework, Foundry, MCP, and Aspire into a working application — an interview simulator where specialized agents hand off control to each other through a defined workflow. The patterns are worth studying: handoff orchestration where each agent fully owns its turn, MCP tool servers that deploy independently of agent code (and can be in a different language), and Aspire tying the service graph together with discovery, health checks, and tracing. The part that stuck with me — the principle of least privilege applied to agent tooling. Each agent only gets the tools it needs. The triage agent gets none. If you're moving multi-agent work toward production, what's the gap you're trying to close? https://lnkd.in/eamzATfP

15

🤖 Think you know what the “R” in RAG stands for? Think again. While most people associate #RAG with Retrieval-Augmented Generation, the retrieval part often hides more complexity than we realize. In this #DataManagement blog, Terry Dorsey, Sr. Data Architect/Evangelist North America at Denodo, dives deep into what retrieval truly means — and why understanding it is critical to building effective #GenAI applications. 🔍 You’ll learn about: - The naïve RAG pattern and its limitations - Retrieval techniques for structured vs unstructured data - When to use each — and why combining them can be a game changer - Emerging new RAG patterns like GraphRAG, Agentic RAG, and Adaptive RAG Whether you’re designing #AI assistants, chatbots, or analytics copilots — this post will help you rethink how retrieval powers generation. 👉 Read “The R in RAG” now: https://okt.to/Y5kciS #LLM #DataRetrieval #DataForALL #NoDataLeftBehind #NoAIWithoutData

20

It’s been an eye-opening journey for our Python and Java SDK teams in Azure Cosmos DB working with OpenAI, as well as all our backend teams helping to meet the scaling needs of the fastest growing application in history: ChatGPT. Scott Guthrie shares insights…

16

LLM serving doesn’t have to be slow or expensive. In this new Domino Blueprints guide, Principal Solution Architect Sameer Wadkar shares how to deploy LLMs as model endpoints—without embedding huge binaries in your model image. Learn how to fine-tune, register, and serve LLMs efficiently using Domino Model Registry and shared datasets. 📘 Stay fast, lean, and compliant: https://gag.gl/Bj70bg #LLMOps #MLOps #ModelDeployment #AI

1

These last 2-3 weeks a lot of folks have been working hard preparing for Microsoft Build. In this journey, due to time constraints, I turned to AI to see how we can get things done faster. There have been many spots where it's saved me time and I'm starting to look at Copilot more as my teammate than a technology. - Finding something: This has been a big help to me. looking back how I used to find things. I'd recall where abouts something was and then my pride would kick in and I'd spend 3-5 minutes going through emails, teams, docs until I found it. Copilot is better at that than me. Sometimes it's a miss, but I'm starting there first now. - Summaries and Actions: Really liking it in meeting contexts. Teams AI meeting generation is pretty good, BUT I caught something and started doing it. I saw people in meetings adding things to chat with actions and items. This helps! just a small tweak, but your future self will thank you. - Demos: WOW. I've gone down a rabbit hole of the concept of a "master prompt" to provide context. I do wish Copilot had some sort of memory capability for me to supply that master prompt. Either I'm not aware, or hopefully it will come. Couple of things I've found on this front. 1. my first prompt is to ask it to generate a prompt for me with a given context. This was HUGE! The detail in just the prompt, and then the outcome that comes from that generated prompt is way more fruitful than just going with a basic prompt yourself! 2. With that generated prompt, AI is actually not bad at creating data to use for demos. (still need to validate you won't get into a problem with the data, but shaves the time down a lot). I'll keep trying things. I'm always working on how to improve productivity and these last two weeks have been a game changer for me.

86

7 Comments

Check out this helpful guide on creating private Snowflake Workspace where you can build and test your code. You can also connect your workspace with a Git repository, offering a powerful, built-in development environment.

4

Every few months, we celebrate a bigger LLM. Impressive—but are we optimizing the soloist when the real unlock is the orchestra? In real workflows, no single model excels at planning, researching, tool use, writing, checking, and editing all at once. Multi‑agent systems divide labor, add redundancy, and arbitrate errors—much like a symphony with sections and a conductor. What’s shifting: - Specialization: a Planner breaks work into steps; Specialists handle retrieval, coding, formatting; a Critic tests and challenges outputs. - Redundancy: multiple reasoning paths vote or reconcile (think ensemble methods for language). - Arbitration: evidence rules and validators catch hallucinations before they reach users. Evidence you can explore: - Microsoft AutoGen (multi‑agent conversations): https://lnkd.in/gW_tyPBm - LangGraph (stateful agent workflows): https://lnkd.in/gPPGXEt3 - Anthropic MCP (interoperability for tools/agents): https://lnkd.in/gsPnsPqk - Self‑Consistency (sample‑then‑select): https://lnkd.in/gUpTmEi3 - Reflexion (feedback‑driven improvement): https://lnkd.in/gKJThjN7 - AgentBench (evaluating LLMs as agents): https://lnkd.in/e8ywbkdN Practical patterns that work: - Support triage: Planner → RAG Researcher → Draft Writer → Fact Checker → Escalation. - Analytics: SQL Agent → Schema Guardian → Data Validator → Explainer. - Marketing ops: Brief Planner → Persona Tuner → Copy Editor → Brand/Legal Checker. - DevOps: Change Summarizer → Risk Checker → Runbook Executor (human‑in‑the‑loop). Design principles: - Name roles and success criteria for each agent. - Require citations: “No evidence, no claim.” - Use budget/latency guards (N samples, max steps, early exit on consensus). - Instrument everything: trace messages, costs, and error reasons. - Keep humans as final approvers where stakes are high. A simple starter: 1) Pick a conductor (LangGraph, AutoGen, or a lightweight message bus). 2) Begin with three roles: Planner, Worker, Critic. 3) Add evidence checkers and retry logic only when it helps quality. 4) Maintain a single‑pass baseline for comparison. 5) Measure win rate, cost, and time-to-result—then iterate. Prediction: the best enterprise AI stacks will be multi‑model and multi‑agent by default. SLAs won’t be about a single model—they’ll be about orchestration quality, observability, and safe escalation. Where have you seen multi‑agent setups beat a single LLM? Which roles would be in your “orchestra”? Share your examples or questions below. Follow for more such content.#AI,#Business,#Technology,#DataScience,#Innovation

1

If you’re using the SQL Analytics Endpoint over Lakehouse data, file health and maintenance really matters. Christopher  Finlan's latest blog does a great job walking through how to ensure your data is in the right shape for fast SQL querying. Check it out https://lnkd.in/gqt4CE4x #MicrosoftFabric #Data #Analytics #SQL #DataEngineering

39

1 Comment

Snowflake's Hybrid Tables: A Frankenstein's Monster with Method to the Madness Like Mary Shelley's creation, Snowflake's Hybrid Tables are an ambitious amalgamation—NoSQL foundations stitched together with relational interfaces. But unlike Victor's tragic experiment, this particular monster serves a purpose. The Architecture Beneath the Hood: Hybrid tables elegantly solve the age-old OLAP vs OLTP dilemma by housing both paradigms in one platform. The underlying NoSQL foundation enables horizontal scaling for transactions, while the columnar layer handles analytical queries. It's like having Jekyll and Hyde coexist peacefully. Finally, Real Constraints for the Relational Purists: For seasoned DBAs who've endured years of Snowflake's "trust but don't verify" approach to data integrity, hybrid tables bring a welcome change. Primary keys are actually enforced at the row level, and foreign key constraints provide real referential integrity. -- Creating a hybrid table with proper constraints CREATE HYBRID TABLE users ( id INT PRIMARY KEY, email VARCHAR(255) UNIQUE NOT NULL, department_id INT, created_at TIMESTAMP DEFAULT CURRENT_TIMESTAMP(), FOREIGN KEY (department_id) REFERENCES departments(id) ); -- This will actually fail if department doesn't exist -- INSERT INTO users VALUES (1, 'john@company.com', 999, CURRENT_TIMESTAMP()); The Performance Reality: There's overhead when queries get rerouted from row-based to columnar execution. Consider this low-cardinality query: -- This query will likely get rerouted to columnar SELECT COUNT(*) FROM users WHERE department_id = 5; In a relational database, this would be a straightforward index scan. In Snowflake's hybrid world, the optimizer might route this to the columnar store, adding latency for what should be a simple lookup. The Verdict: The engineering is genuinely clever - they've created a unified platform that handles both transactional and analytical workloads while providing referential integrity. The performance trade-offs are real but manageable. Sometimes the best solutions aren't the purest ones - they're the ones that actually solve real problems. #Snowflake #DataEngineering #DatabaseArchitecture #ReferentialIntegrity #OLTP

26

5 Comments

LLM serving doesn’t have to be slow or expensive. In this new Domino Blueprints guide, Principal Solution Architect Sameer Wadkar shares how to deploy LLMs as model endpoints—without embedding huge binaries in your model image. Learn how to fine-tune, register, and serve LLMs efficiently using Domino Model Registry and shared datasets. 📘 Stay fast, lean, and compliant: https://gag.gl/Bj70bg #LLMOps #MLOps #ModelDeployment #AI

1