AI & LLM

If you’ve been waiting for a computer-use agent you can run on your own machine (sandboxed, human-in-the-loop, open source), Microsoft’s Magentic-UI is the repo to clone this week. It now leads with MagenticLite (“Big tasks. Small models.”), a redesign from Microsoft Research AI Frontiers that pairs a small-model orchestrator with a dedicated browser-use model and keeps you in control the whole way.

The OWASP LLM Top 10 exists because shipping an LLM to production without a security model is a new category of risk that the existing web application security playbook doesn’t fully cover. Prompt injection has held the #1 spot on that list since the first version was published, and it’s not there because researchers think it might be a problem someday. It’s been demonstrated against production systems at companies that knew what they were doing.

Before MCP existed, adding tools to an AI application meant writing the same glue code over and over. You had OpenAI’s function calling syntax. Anthropic had tool use with a slightly different schema. LangChain abstracted over both, but now you depended on LangChain’s versioning decisions. Every new model provider meant rewriting your tool definitions. Every new tool meant re-registering it across every AI integration you maintained.

Every vendor selling software right now claims their product is “agentic.” I’ve seen chatbots with a system prompt called an agent. I’ve seen a scheduled Python script described as autonomous AI. I’ve also shipped actual agents to production, at an insurance company, handling FNOL triage, policy lookup, and claims routing. The gap between what gets marketed as an agent and what you’d actually build in production is wide, and it’s expensive to get wrong.

I was at Microsoft Ignite for the Foundry announcements, and the framing runs ahead of the substance. Microsoft rebranded most of its Azure AI stack as Microsoft Foundry: one PaaS spanning models, agents, tools, knowledge, and governance. Underneath the launch-day gloss, three things actually matter to anyone shipping enterprise agents.

Docker shipped a set of tools that let a platform team decide exactly which Model Context Protocol servers developers can run: a catalog format, an open-source Gateway, and a Toolkit built into Docker Desktop. The official catalog carries 220+ containerized MCP servers, and you can fork it, trim it, or build a private catalog that points at images in your own registry.

At some point I had three side projects each talking to a different LLM provider, with API keys pasted into three .env files and three slightly different client wrappers. Swapping GPT-4o for Claude in any of them meant editing code. That is a silly amount of friction for what is, underneath, the same chat-completion call, so I put a LiteLLM proxy in front of everything and never went back.

The first time I let Copilot loose on one of my Angular repos, it wrote a perfectly reasonable component for 2021: constructor injection, class properties for state, NgModule imports. We were on Angular 19 with signals and standalone components.

TL;DR: This guide walks you through building a production-ready RAG system using FastAPI, ChromaDB, MinIO, and OpenAI. Learn document chunking, vector embeddings, hybrid search, and real-world deployment strategies. Introduction # As a .NET developer watching the AI space move fast, I found myself both excited and skeptical. When tools like Claude.ai and ChatGPT started offering out-of-the-box RAG solutions, I wanted to build my own system with full control over the implementation.

I run LLMs on my own machine for two reasons: my prompts never leave my hardware, and I can experiment all day without watching an API meter. The setup used to be the hard part. Wrangling Python environments, CUDA drivers, and model weights by hand was enough friction that most people quit before the first token.

Business users kept asking my team for one-off data pulls: how many orders above $150 in December, which products are low on stock, that sort of thing. Every request meant a developer dropping their work to write a throwaway SQL query. So I built a REST API that does the translation instead. It feeds the database schema to an LLM, gets back a SQL query, runs it, and returns JSON. One codebase, four interchangeable providers (OpenAI, Azure OpenAI, Claude, and Gemini), switched by a single environment variable.

Introduction # I built this chat app to answer a narrow question: how little infrastructure do you actually need to stream tokens from Azure OpenAI into a browser? Less than most tutorials assume. There is no SignalR here and no WebSocket hub.