AI Experiments

I stopped researching which expensive translation tool to buy and built an orchestrated AI agent instead — inside Microsoft 365, using Copilot Studio, without leaving the tenant.

I want to build a live AI assistant that can hear a user and see their screen at the same time, then talk back in real time. This post walks through getting Google’s ADK bidirectional streaming demo running locally with mic, camera, and voice out — and documents the things that tripped me up.

In the same week that OpenAI and Anthropic dominated AI headlines over a rushed military deal, Google was quietly shipping the protocols that will define how agents browse, buy, and collaborate on the web: WebMCP, the Universal Commerce Protocol, and Agent2Agent.

I tried to recreate a Google AI Studio proof-of-concept in code. The first path failed on authentication; the second produced a working bidirectional streaming prototype; the third pushed it toward a realistic scenario.