Friday night: make one lamp answer to you
Voice in, relay click out, and the millisecond count that says where the wait actually lives.
What you get out of this
By the end of the evening you say "lights on" and a lamp comes on. That is the easy half, and plenty of tutorials stop there. The other half is the one worth having: a millisecond number for each leg of the pipeline, measured on your kit and your network, and a second number after you move one leg off the internet.
The first version will feel sluggish. That is on purpose — the whole build is about finding out which part is sluggish rather than assuming.
Kit
- A Wi-Fi smart plug with a local HTTP API. Shelly and Tasmota-flashed plugs both have one. Around £10–15. A plug that only works through the vendor's app will not do — you would be measuring their cloud, not yours.
- A lamp. Any lamp. Nothing that heats: no kettles, heaters, or hair dryers.
- A microphone. The one in your laptop is fine.
- No smart plug? Do the whole build anyway with a printed line and a terminal bell in place of the relay. Every measurement in here still works; you just lose the click.
The shape of it
Three lessons, about ninety minutes, all free. You are done when the lamp responds and you have a before-and-after latency split written down.
Friday night
Working, then measured, then faster — in that order.
- Get it clicking, badly — The whole loop, end to end, in the ugliest form that works. Speed comes later. (25 min)
- Find the second that went missing — Split the pipeline into legs and find out which one owns your latency. Most people are wrong about this. (30 min)
- Cut the round trip — Move transcription onto the machine that is already listening, and measure what that actually bought you. (30 min)
Related writing
- How to measure LLM latency: find which leg owns the two seconds — A voice command that takes two seconds to fire has four legs and only one of them is slow. Here is how to time each one separately before you optimise anything.
- From a token to a servo: the last ten centimetres — Everything between a model's output and something physically moving — the boundary, the failsafe, and why the interesting engineering is all on the hardware side of the API call.