Introduction: Natural Language as the New Programming Interface

English as a programming language. Historically, you would need to use assembly language, Python, Rust, etc. But these days, with LLM agents, you can just go to Codex and ask him to do something and he'll write the code really quickly.

Project Overview: A Multimodal AI Device

So, I built an AI device that can see, listen, speak and think.

So with its camera box which you have a button which you press to speak and there's also a different version where you press to wake and with the camera you can take a picture also you can play an interactive game.

So inside the box we have a Raspberry Pi, Raspberry Pi 3, a voice, hardware on top, a microphone strip, an LED button and a USB webcam.

Hardware Setup

System Architecture: Live Loop, Tools, and Task Handoff

So this is a live session loop. So we have the wake, the listen, the The C, the tools, such as taking a picture, playing music, etc.

And it can speak, and then it will close. And it either closes when you tell it to close, or when you don't speak with it for a long time.

So this is an async task hand -off. so we have the real -time model which speaks live with you then we have the task queue so as we saw here the tools so if you ask it to play music it will put it to the end of the task queue and then a GPT 5 .5 worker will take it off

so it can either search the web execute the code or find the weather and then it wakes the bot when it's finished and then it updates.

Second Build: An AI-Powered Watch

So now I will be presenting my watch which I also built using AI.

So inside the watch it has ESP32 S3 with 240 megahertz, megahertz Wi -Fi and the Bluetooth low energy a touchscreen which is fairly little a power chip and a motion sensor and usually this would last for two

hours maximum now I can wear it for like two months and nothing will happen Because unless I put my hand over it to turn it on, it's in deep sleep, which uses very, very little energy.

So here's the runtime loop. So I put my hand over to turn it on. Then I choose what I want to do so I can change the brightness and I can turn on and off the voice.

I can also fetch things by pressing up here it will use Wi -Fi it will fetch the new it will update the weather instantly normally it updates once every 10

minutes because connecting to the Wi -Fi uses a lot of energy and then it sleeps deeps, goes into deep sleep to save energy.

Runtime and Power-Saving Loop

Weather + TTS Call Sequence

So this is the call sequence.

You turn on the Wi -Fi for a little bit because else it uses a lot of energy, then it generates a string, takes the weather, then it merges the two strings and sends it to TTS from OpenAI

and then it saves the audio and plays on the speaker locally on the watch so this

Building with Codex: The Agentic Development Loop

is a Codex build loop and I use Codex to make all of these devices so I use slash goal to make it so that I can set it up in the morning and then when I come back from school using the goal, it will plan, then edit, then test, deploy and debug and it will keep doing that and it will make its own prompts.

And then this is what Codex can do today.

So I used 251 million tokens for the voice bot. I calculated this by asking in Codex to find out how many tokens I used, and the watch uses 67 million tokens.

I used GPT 5 .5 extra high fast, and I used about 2 ,440 terminal cores, and I did way less prompts.

fonts.

Evaluating Models: DeepSui Software Engineering Benchmarks

Now I will be showing DeepSui software engineering, not puzzles. It has 113 long horizon tasks, behaviour verifiers, 91 real repositories, so on.

This is a DeepSui leaderboard.

So, as you can see, Clod is on the top, then goes GPT 5 .5, which is just a bit worse, but costs three times less, and then in third place we have Opus, which costs two times more than the VESA model, and this is the end.

Demo and Build Process

And now I will be showing my demo. So I have the demo here. Due to Wi -Fi problems, this doesn't work.

And so here are some pictures of me building it.

Repurposing Old Hardware

So I took an old Google Voice app, which was made long before AI, and it was somewhere in the attic. And it was very dusty.

Assembly, Soldering, and Final Notes

So as you can see I was fixing it, I was also soldering it, as you can see the soldering iron, I was re -soldering the wires and it's actually surprising that old hardware with AI can build such a modern device.