AI x Mechatronics Engineering New Grad (Part-time)

Vancouver, Canada · In-person, part-time (2–3 days/week) · 4-month term, extendable · Paid: $30-$40/hourly

THE SHORT VERSION

Every AI demo you've seen lives on a screen. The model reads text, writes text, and stops at the edge of the display. I want to know what it does when it gets a camera, a microphone, a motor, and a reason to act. Almost nobody is testing that seriously with cheap hardware, and that is the whole job.

I run DigitalSamaritan, 1M+ followers across channels, built on testing AI claims and reporting what actually happens. I'm setting up PracticalyAI Labs, a hardware lab in my studio, and I'm hiring one mechatronics intern to run it with me. The Labs has three jobs:

1. Push the models. Take a frontier model out of the chat window and find where it actually breaks when it has to see, hear, and act in the real world. The real limit, not the marketing one.

2. Hack physical things. Open up devices that already exist and give them a brain. A thermostat, a doorbell, a toy, a lamp, an old phone, a kitchen scale. Find the pins, tap the serial line, flash it, and put a model in the loop.

3. Prototype gadgets. Build objects from scratch that only make sense now that a model can see and act, and ship one every week.

Every build goes in front of a large audience, wins and failures both.

WHAT WE'RE ACTUALLY TESTING

Each build starts with a capability question. The best ones also carry a question about daily life: what stops needing a phone, what you'd stop typing, what you'd actually let a motor touch. Examples of the kind of thing on the list:

• Can a vision model on a $60 board watch a room and act on what it sees, reliably enough to trust?

• How small can a local language model get before a device is useless without the cloud?

• Can a device hold a spoken conversation, with tool calls, at a latency that doesn't feel broken?

• What does an agent do when its output is a motor, not a paragraph? Where does it fail first?

• Take a $20 device off a shelf. Once you're on its serial line, how much of it can a model take over?

• Can one person plus a model design, print, and ship a working device in a week?

Some of these will work. Some will fail on camera. Both are the content.

WHAT YOU'D DO

• Ship every week. One working prototype per week. Ugly is fine. Breadboards and hot glue are fine. Working is the only requirement.

• Build the prototypes. Firmware, sensor wiring, power, enclosures. ESP32, Raspberry Pi, Jetson-class boards, e-ink and small displays, cameras, microphones, speakers, motors, whatever the experiment needs.

• Tear into existing hardware. Open the thing up, find the test pads, sniff the bus, dump the firmware, and wire a model into whatever it was already doing. A multimeter and a logic analyzer are your main tools.

• Put the model on the device. Cloud APIs when latency allows, local and quantized models when it doesn't.

• Find the limit. Push every build until the model fails, then write down exactly where and why. That number is the result, not the demo.

• Give it a body. Basic CAD and 3D printing so the thing is an object someone would keep on their desk, not a breadboard on a napkin.

• Instrument everything. Latency, failure rate, battery life, how often the model was wrong. If we claim something works, we show the numbers.

• Pick next week's build. Weekly idea session. You bring what's buildable in five days and what you'd want to crack open next. We pick the one most likely to teach us something.

• Document as you go. Photos, clips, and a build log for every project. That's the raw material for the video and it's your public portfolio.

WHO THIS FITS

• You're a mechatronics, electrical, or mechanical engineering student or recent grad, and you've built things nobody assigned you.

• You've opened up something that wasn't meant to be opened and got it to do something it wasn't meant to do.

• You can show a microcontroller project that worked, with proof, and explain what broke along the way.

• You can solder, read a datasheet, and debug a circuit that worked yesterday and doesn't today.

• You've written Python or C/C++ for hardware and you've called an API from it. Bonus if you've run a model on an edge device.

• You use frontier AI tools daily, for code, for CAD, for debugging, and you have opinions about where they fall over.

• You have opinions about what these models should and shouldn't be allowed to touch, and you'd rather build the test than write the essay.

• You're comfortable with jank. An 80% prototype this week beats a perfect one next month.

Nice to have: PCB layout (KiCad), running quantized models on edge hardware, UART/I2C/SPI sniffing with a logic analyzer, custom firmware on consumer devices (ESPHome, Tasmota, OpenWrt, or your own), ROS or basic robotics, any audio or camera pipeline work on embedded boards, and a willingness to be on camera occasionally.

WHAT YOU GET

• Paid, in person, with a real budget for parts, boards, devices to tear apart, and API credits.

• A public portfolio of working builds with your name on them, seen by a large audience. At one a week, that's a lot of portfolio.

• Hands-on time with the newest models the week they ship, on hardware, not in a chat window.

• A reference from someone who watched you take an idea to a working object under a deadline, every week.

HOW TO APPLY

Send:

1. Two things you built or hacked, with a photo or short video of each and one paragraph on what broke and how you fixed it.

2. One experiment you'd run in a weekend with a $150 parts budget that puts a model on a device, or inside a device that already exists. State the question about the model, the parts, where you expect it to break, and how you'd know if it failed.

3. Your availability and where you're based.

• Applications without #2 won't be read. That exercise is the job.