Live links + architecture write-ups for demos built on Azure. Each demo explains the services used, data flows, and the design decisions and trade-offs so an engineer can understand and rebuild it.
A private kids' storybook service built with Azure OpenAI gpt-5.4 (story) + gpt-image-1 (art) + speech synthesis (TTS). The child's photo is NOT stored — it's used only transiently to derive a character description (data minimization, KR PIPA-aware); page art is generated in parallel to cut wait time; and everything is persisted on Container Apps + PostgreSQL (keyless MI). It explains the design decisions and trade-offs — billing limits, per-user isolation, and governance avoidance. (Private — no link.)
View architecture →An investment-analysis demo built with Azure OpenAI gpt-5.4 (keyless) + function calling (tools) + multi-source real data (SEC · FRED · OpenDART · ECOS). Value/technical/risk/chair roles debate, but every number is fetched by the agent via a tool and cited (hallucination prevention). It explains the design decisions — keyless security, real-vs-illustrative source honesty, and $0 hosting.
View architecture →An auto-content portal built with Static Web Apps + Container Apps Job + Azure OpenAI (keyless) + Storage ($web history). It walks through the architecture and design decisions — social login, custom domain, and unified static/dynamic section deployment — so an engineer can rebuild it.
View architecture →An automated shorts demo built with Foundry gpt-5.4 (5 logical agents) + Sora 2 + Service Bus (Premium) + a Container Apps worker (KEDA) + Blob. It explains the design decisions and trade-offs — human approval before paid calls, Private Endpoint isolation, keyless Managed Identity, and idempotent resume (avoiding double billing).
View architecture →A case study where central governance periodically forces Storage publicNetworkAccess to Disabled: the Container Apps Job is moved into a VNet-injected env and the storage gets a Private Endpoint (blob) + Private DNS, so it keeps running while coexisting with the policy. It explains the design decisions and trade-offs — env recreation, UAMI reuse, and VNet peering + cross-registered DNS for the two jobs' cross-mirroring.
View architecture →A book-generation platform built with Azure OpenAI gpt-5.4 (keyless) + Container Apps + WeasyPrint. It raises quality through 8 stages (plan→grounding→outline→write→self-critique·revise→consistency→copyedit→typeset) and blocks hallucination via evidence grounding from references (files·URLs·web). It explains the design decisions and pitfalls (Korean fonts, header encoding, governance storage) for producing a typeset PDF with cover, auto TOC, body, and bibliography.
View architecture →