Skip to the content
2 layers100 × 79 mmGpt alarm, the printed circuit board in 3D

Vision Alarm Clock

Size
100 × 79 mm
Copper layers
2
Thickness
1.6 mm
Parts
18
Nets
16
Pads
48
Tracks, vias
59, 5
Pours
2
Routed
Yes
KiCad
KiCad 10.0 (file 20260206)
Created
12 Apr 2026
Updated
28 May 2026

Overview

Gpt alarm is an open-source printed circuit board by kyle-tennison, from the kyle-tennison/night-vision repository on GitHub. It is a 2-layer board of 100 × 79 mm with 18 parts on 16 nets, routed with 59 tracks and 5 vias.

It is built around the 1N4148WS and AO3400A. Its parts include 8 connectors, 3 resistors, 3 diodes, 2 transistors and 1 switch.

Main components

2

README

README.md

Overview

Getting up in the morning sounds simple, but for many people, the hardest part is staying out of bed after the alarm goes off. Traditional alarms only solve half the problem: they can wake you up, but they can’t stop you from crawling back under the covers five minutes later.

Night Vision takes a different approach; it's an AI-powered alarm clock that uses an infrared camera and a fully local AI model to determine whether you’re still in bed. The alarm goes off at your set time, and only turns off when it sees you get out of bed. If the system detects that you’ve gone back to bed, the alarm starts again automatically. Once you’re actually up, it stays silent.

Privacy

Having a camera pointed at your bed is unnerving to most people. However, it's really no different than having your phone camera near your bed. Nothing is being sent over the web: the photos are streamed directly to an on-prem vision-language model, and are never even written to disk.

Technical Overview

This project is implemented in Rust on a Jetson Orin Nano dev board. This is more of an embedded project than an AI project. Fast iteration is desired, so the following system was chosen in an attempt to maximize efficiency.

GStreamer

The gstreamer-rs crate was used to implement GStreamer directly in Rust. This allows frames to be read directly from a callback function. This bypasses any network or file I/O, and doesn't need to make any kernel calls whatsoever because everything happens in shared memory1.

There is a bit of time lost between requesting a frame in the "Frame Request Mutex" and the time the callback is triggered. However, the current implementation prefers a more recent frame over the fastest frame. There is no point in running the loop on an old frame.

Vision Language Model

llama.cpp is used for all AI-related processing. It was recompiled to take full advantage of the NVIDIA hardware. The model of choice is a tiny 450M vision-language model from LiquidAI: LiquidAI/LFM2.5-VL-450M-GGUF. With the NVIDIA chip, the entire loop, from image to prediction, takes 0.3s.

CAD Model

A public Onshape CAD model is provided here. I was going for an 80s beige vibe, which a lot of people hate, but I personally love.

PCB

A very simple PCB sits under the Jetson Orin Nano dev board. KiCad files for this are in board/. There is a spot for a relay here that is unused, but might be useful in a future version. The main purpose of this board is to operate the buzzer and an indicator LED.

Unrelated

15-year-old Attempt

For a very long time, I've wanted an alarm clock that could detect when I was in bed. When I was 15, I tried a very similar approach using a very sketchy force transducer made from my parents' bathroom scale. This is what it looked like:

image

This worked for a couple of days, but I had to plug my laptop into it every night, so it didn't last long. I remember at the time contemplating running a CNN on a Raspberry Pi that could monitor a video feed of my bed. I actually found the blog post where I was thinking about this:

"I also considered the option of training an AI to use a picture to determine if I was in bed or not. I also ended up deciding against this because of the sheer complexity."

From a blog post in 2020.

And it's good that I dropped that idea, because a conventional CNN would be extremely sensitive to any change in the environment. That is if I could even get it trained in the first place. It would require a lot of training data, and the moment my room changed (new picture, different bedding, etc.), the model could easily start spitting out garbage.

Footnotes

  1. Don't quote me on the kernel part—I might be mistaken. ↩

Parts

18
QtyValueFootprintReferences
210kResistor_SMD:R_1206_3216MetricR1, R3
21N4148WSDiode_SMD:D_SOD-323D3, D4
2AO3400APackage_TO_SOT_SMD:SOT-23Q1, Q2
1100Resistor_SMD:R_1206_3216MetricR2
119V INConnector_BarrelJack:BarrelJack_HorizontalJ3
119V OUTTerminalBlock:TerminalBlock_Xinya_XY308-2.54-2P_1x02_P2.54mm_HorizontalJ4
1BUCK IN +Connector_PinHeader_2.54mm:PinHeader_1x02_P2.54mm_VerticalJ6
1BUCK IN -Connector_PinHeader_2.54mm:PinHeader_1x02_P2.54mm_VerticalJ5
1BUCK OUT +Connector_PinHeader_2.54mm:PinHeader_1x02_P2.54mm_VerticalJ8
1BUCK OUT -Connector_PinHeader_2.54mm:PinHeader_1x02_P2.54mm_VerticalJ7
1BuzzerBuzzer_Beeper:MagneticBuzzer_StarMicronics_TMBBZ1
1Conn_01x05_PinConnector_PinHeader_2.54mm:PinHeader_1x05_P2.54mm_HorizontalJ1
1HK19F-DCxxV-SHCRelay_THT:Relay_DPDT_Finder_30.22K1
1LEDDiode_SMD:D_1206_3216MetricD1
1Siren_OutTerminalBlock:TerminalBlock_Xinya_XY308-2.54-2P_1x02_P2.54mm_HorizontalJ2

Similar boards

  • 4 layers140 × 100 mmDc usv, the printed circuit board in 3D

    Dc usv

    TobleMiner/DC-UPS

    A small DC UPS, designed to keep small appliances like consumer network equipment running during power outages

    351 parts223CERN-OHL-P-2.0
  • 4 layers118 × 117 mmCm4-robot, the printed circuit board in 3D

    Cm4-robot

    Twisted-Fields/acorn-robot-electronics

    PCB hardware designs in KICAD for Acorn, the precision farming rover from Twisted Fields.

    268 parts81
  • 4 layers55 × 100.5 mmOSS Radio Hardware Design, the printed circuit board in 3D

    OSS Radio Hardware Design

    Arkos-Engineering/HT-15_hardware

    We can split this out into separate repos later, right now its very convenient to have everything here.

    256 parts72
  • 4 layers41.4 × 108 mmMesh3310, the printed circuit board in 3D

    Mesh3310

    JLTelectronics/The-Mesh3310-Project

    Make the worlds strongest Node

    223 parts26MIT
  • 2 layersTopBoard, the printed circuit board in 3D

    TopBoard

    Libre-Talk/Wave

    Smart Speaker satellite

    184 parts22
  • 4 layers115.5 × 96.2 mmClosed deck x4, the printed circuit board in 3D

    Closed deck x4

    OpenLogicEFI/Hardware

    Hardware is based off an open source project called Speeduino by Josh Stewart.

    245 parts19
  • 2 layers40 × 20 mmDutyCycle, the printed circuit board in 3D

    DutyCycle

    FASTSHIFT/DutyCycle

    Work-Life Duty Cycle Clock

    46 parts18MIT
  • 4 layers100 × 80.3 mmHPM6P41 BB, the printed circuit board in 3D

    HPM6P41 BB

    Alipay1/HPM6P41_BB

    以HPM6P41为主控的电源开发板

    334 parts14MIT

Questions

How big is the Gpt alarm PCB?

It measures 100 × 79 mm and has 2 copper layers, 1.6 mm thick.

What parts does the Gpt alarm use?

18 parts, built around the 1N4148WS and AO3400A. The full list of its parts is on this page, with links to each part's symbol and footprint where PCBKit's library has them.

Where can I download the Gpt alarm KiCad files?

From the kyle-tennison/night-vision repository on GitHub, at the commit this page was made from: the board is board/gpt-alarm.kicad_pcb and its schematic board/gpt-alarm.kicad_sch. PCBKit opens the board in the browser, with nothing to install.

Can I use the Gpt alarm design in my own project?

Its repository states no licence, so its author keeps all rights: ask kyle-tennison before reusing the design.