Hardware & Electronics

A $2 web server that can think

An ESP8266 chip running a web server on a 9V battery. Back then it switched a light. Today it can ask a language model what to do next.

A NodeMCU ESP8266 board powered by a 9V battery
Type
Experiment
Status
Completed
Year
2015
Focus
ESP8266, IoT, Edge AI, Wi-Fi, LLM

The chip costs a few cents. The board around it costs about $2. It has Wi-Fi, a 32-bit processor running at 80 MHz, and enough memory to serve web pages. Five times the clock speed of an Arduino UNO, in a module smaller than a postage stamp.

That's the ESP8266, and it changed how I think about "smart" devices.

The tiny ESP8266 module that started it
The tiny ESP8266 module that started it

A web server in your pocket

Back in 2015 I wired a NodeMCU board to a 9V battery and an LED, and loaded a small sketch. The board joined the Wi-Fi network, started an HTTP server, and served a page with two buttons. Tap On from any phone on the network and the light came on.

No hub. No cloud. No app store.

The page served straight from the chip
The page served straight from the chip

Nice demo. But the important part was the idea underneath: any physical object can have its own server. A lamp, a fridge, a door, a plant pot. Each one can expose a little API, and each one can talk to the internet.

I used the same trick for a desk lamp that lights up when you get a notification.

The IKEA social notification desk lamp
The IKEA social notification desk lamp

Then the internet got a brain

If a device can call an API, it can call a language model.

That flips the design. Instead of hard-coding every rule (if temperature > 24, open vent), the device describes what it sees and asks for a decision. A small, fast, cheap model is perfect here. It doesn't need to write poetry. It needs to read a handful of numbers, understand context, and reply with a clear action in JSON.

The pattern is simple:

  1. Sense. The chip reads its sensors.
  2. Describe. It sends a short, structured summary to the model, along with the time, the season, maybe a forecast.
  3. Decide. The model returns an action and a reason.
  4. Act. The chip flips a relay, moves a motor, sends an alert.

Rules break the moment the world does something you didn't plan for. A model can reason about the odd case.

A few things I'd love to build

A greenhouse that reads the weather. Soil moisture is low, but rain is forecast in two hours. A rule waters anyway. The model waits, saves water, and tells you why.

A shop fridge that notices patterns. Door-open counts, temperature drift, time of day. It spots that the door keeps getting left ajar after the lunchtime rush and flags it before stock spoils.

A care sensor that knows a normal morning. Kettle on, bathroom light, front door. When the routine shifts in a way that matters, a relative gets a calm, specific message. Ordinary variation doesn't set off an alarm.

A vending or coffee machine that diagnoses itself. It reads its own error states, works out the likely cause, and books the right engineer with the right part.

Each one runs on hardware costing a few pounds, with a model call costing a fraction of a penny.

What I bring

This sits right where my disciplines overlap: electronics, embedded firmware, cloud APIs and AI. I can prototype the device, design custom boards when off-the-shelf won't do, and build the decision layer and the dashboard that sits on top.

Got an object that should be smarter? Tell me about it.

More on this project

Have a project in mind?

Let's build