App Modernization Done. Now Let's Make It Smarter.

TL;DR

• 1

  Semantic search fixes obvious UX dead ends: In Bruno's eShop Lite demo, a keyword search for "cooking" finds nothing, while semantic search returns camping cookware and a camping stove, and "rainy day" surfaces a rain jacket plus a survival kit.

• 2

  The implementation is lighter than the jargon suggests: The sample uses SQL Server for products, an in-memory vector store for the demo, Microsoft.Extensions.AI libraries, embeddings per product, top-3 vector search, and a simple score threshold above 30 percent before GPT-5 mini writes the final response.

• 3

  Observability can become an AI feature, not just a dashboard: Bruno adds a local observability assistant that analyzes the last 10 minutes of logs with Foundry Local, clusters failures, and spots injected HTTP 500 errors from both English and Spanish queries.

• 4

  Small local models are good enough for useful app intelligence: His local setup uses Phi 3.5 mini for chat, an all-MiniLM embedding model around 80 MB, and Foundry Local with GPU, CPU, OpenVINO, and ONNX-compatible options to avoid sending every analysis task to the cloud.

• 5

  Business signals matter more than raw traces: In scenario 15, a separate Store Intelligence service groups search behavior like hiking and rain questions, then generates a report showing product gaps, top customer intents, and no-result searches such as repeated requests for hiking boots.

• 6

  MCP and agents are the path from smart feature to composable system: Bruno shows how catalog, promotion, observability, and business insight services can be exposed as tools or agents, orchestrated with Microsoft Agent Framework, and even published through Foundry using agent-to-agent endpoints and agent cards.