Local AI Agent: Wikipedia Vandalism Detection

"Streaming" in AI agents has two meanings. It can mean an agent consumes a live stream of events or that its output appears token by token. A new build demonstrates both, solving different problems for a truly useful, always-on agent. This specific build creates a local agent that monitors Wikipedia's live public edit feed. It aims to identify potential vandalism, running entirely on your own machine using Ollama. This project requires Python 3.11 or newer and a local Ollama installation with a pulled model. No API keys or cloud accounts are needed, and there are no costs beyond electricity. The key design decision involves a two-stage filtering process. The first stage uses basic Python to quickly filter out uninteresting edits. Only a small fraction of events that pass this initial filter then trigger the more resource-intensive local language model. This approach prevents the agent from falling behind the live stream and conserves computing power. This matters because it shows how to build efficient, always-on AI agents without relying on external services or high costs.