NeuroAR

NeuroAR are AR glasses that read brain activity. It is a pair of glasses with a small OLED display reflected through a lens, prism, and mirror-acrylic optical path.

It has forehead electrodes connected to a custom board using the XIAO ESP32-S3 and ADS1292 biopotential front-end. The basic flow is body signal to custom board, custom board to XIAO, and XIAO to display feedback.

The final simple demo includes reading the brain activity and showing feedback through the wearable display while being connected to an external dashboard/web app to control the screen.

More advanced spirals include processing the EEG signals using SNNs, connecting NeuroAR to Poke through MCP, mounting this technology and connecting it with Meta Ray-Ban Glasses, and making a system that uses the features across what I developed and what's on the glasses.

One of the biggest inspirations for the display side was Mañolo's YouTube channel. One of the initial videos I saw while thinking about my glasses was this one. It showed a simple AR display system that I scaled up a bit for my own glasses.

I also looked at projects close to mine so I was not designing in a vacuum. Open Source Smart Glasses is useful for open smart-glasses architecture; it was a really helpful starting point for my ideation process. That project eventually became the basis of a YC company called Mentra (I'm in contact with their founders); they are building an open-source OS for smart glasses. Uwear was another project I used for inspiration on the optics.

On the EEG side, the e-Glass paper shows EEG built into an eyeglasses form factor, with frontotemporal electrodes, real-time monitoring, and edge-ML applications like seizure detection and cognitive workload monitoring. Open-Source EEG on a custom PCB and OpenBCI were good references while building the EEG system, and I went deeper into the TI ADS1292 datasheet.

For tools, I used ChatGPT, Codex, and Claude for brainstorming, code help, debugging, and documentation cleanup (only cleanup, the actual documentation was written by me, as I explain in How I used AI in Fab Academy). The actual design and build tools were KiCad, Fusion 360, Inkscape, PlatformIO, and Arduino C++, with MicroPython for the ML processing spiral.

Starting from the CAD design process: I used both 2D and 3D design, from the project's logo to the glasses frame, the side housing, the temples, the display holder, the lens/prism position and optics, the mirror-acrylic combiner angle, the cable concealment, and the electrode placement.

I used 3D printing (additive manufacturing) to produce the main frame and enclosure for all the components, and laser and vinyl cutting for both the mirror combiner component and the decorative final touches (subtractive manufacturing).

Moving to the electronics design and production: I designed all the electronics and their wiring, connecting and controlling the display, the ADS1292, the XIAO ESP32-S3, and the system for reading the recorded brain activity, and interacting with the user through both software and the input devices (EEG electrodes and button), in KiCad, and produced it on an FR4 copper board using our PCB milling machine.

Now to the software production: I designed a system to interface with my AR glasses over USB and WiFi while enabling the user to use the input devices to interact with the screen, and I connected the screen with external services that give the glasses more functionality including email notifications. I also designed the full software and dashboard plus the ML to process signals from the glasses, with the help and "mentorship" of Codex and Claude.

The glasses were fully integrated, ensuring everything fits and snaps in place without any wires being visible, with the display module detachable and the electrodes sticking to the forehead. I made sure the system can be worn to a comfortable extent with power reaching it from a customized small powerbank.

The main parts are the XIAO ESP32-S3 Sense ($13.99), the ADS1292 ($9.84), the SSD1331 OLED (~$10), the plano-convex lens ($31), the beam-splitter prism ($10), the electrodes, the PLA, the mirror acrylic, the vinyl, and the passives for the board.

Most parts came from DigiKey, Amazon, Akihabara, AliExpress, and lab stock, and the minimum total is about $127.78 (excluding shipping and fees). A lot of the materials were already available at the lab, and I ordered more components than this for different experiments like a higher quality camera and different electrodes and connectors, but these are the parts that are actually in the project.

The full breakdown, with every component, its source link, and its cost, is on the Bill of Materials page.

The biggest one was whether this whole thing could actually live on a pair of glasses instead of staying as boards on a desk, as in whether I could combine AR and EEG signal processing in one form factor. That one got answered: the XIAO, the ADS1292 board, the wiring, and the display optics all fit into the temple and the detachable module, everything snaps in, and you don't see any wires from the outside (the electrodes just stick to your forehead).

A question I set early was whether I could reuse the lenses from my own glasses instead of making custom ones. Cutting or ordering custom lenses is expensive and a whole separate process, so I designed the frame from the outline of my actual lenses, and they dropped straight in.

I was curious whether I could make the design all snap together without any screws. I learned how to design proper snap-fits, and the temple cover, the display module, the prism holder, and the wire concealers all click into place, which also keeps the cables hidden.

The hinge was very annoying too. My first idea of printing the hinge in one piece with the frame failed instantly, it snapped the moment I moved the temple. After a few iterations and a tutorial, I got a separate hinge that folds and holds in place.

Balance was something I wasn't sure I could get right either. Splitting the electronics across the frame, the EEG chip near the temple electrodes and the XIAO toward the back, and routing the powerbank cable out the back, ended up balancing it well enough to wear.

I also wanted to know if I could read a real biopotential signal off a board I made myself. For strong forehead events like blinks and eye movement, yes, the ADS1292 picks them up.

The optics question got answered too. With just the screen and the prism the display had to sit way too far from my eye, but adding the convex lens roughly halved that distance, and the angled mirror-acrylic combiner only reflects the light from the prism into my eye while leaving my forward view mostly clear.

There was also a fabrication question: whether our PCB mill could even handle the fine-pitch ADS1292 in its 32-TQFP package with 0.2 mm traces. After ruining the first board and the chip by being impatient with the solder, it milled and came together cleanly with the right settings.

The last practical one was power. The glasses run off a small external powerbank over USB-C from the back of the right temple, and that also helps balance the weight, so I could actually wear it for a reasonable time.

The integration and packaging are what worked best. I designed the frame from my own lenses, turned the right temple into the electronics enclosure, hid the cables, and snapped everything together with no screws. The result is a prototype that actually looks like one designed object instead of parts taped together. I wouldn't necessarily want to daily-drive the glasses, but it is clearly a working prototype.

Optics worked too. The OLED, lens, prism, and mirror combiner line up at a good distance, the display is readable, and it doesn't block my normal vision when I'm not looking at it.

My milled ADS1292 board did its job. It powers up, talks over SPI, and reads strong forehead biopotential signals like blinks and eye movement.

Software held up through all the hardware changes. The small front-end feeding the XIAO, the XIAO as the hub, the browser dashboard, the recording and training pipeline, and the board-native Display Studio all worked together.

I'm still actively experimenting with the SNN side, and it's kind of working. Trained on my own recordings, it pushes strong-event detection above a naive baseline, so it already adds something on top of raw thresholding. It's not a finished, reliable controller yet, mostly because my dataset is tiny, but the pipeline runs end to end and the direction looks promising.

The advanced EEG is the part that didn't get there. Clean, separated brainwave bands and any real mind-state reading aren't happening yet, and that needs a fundamental hardware change: more electrodes and an ADS1299, which is only really possible through a board house. The raw signal is still too noisy, and that points at physical things first: only two electrodes, no dedicated reference/bias electrode this spiral, plus motion, contact, and cable noise.

My earlier DIY EEG front-end didn't really work well either. The op-amp plus SAMD11 board I built to read the signal was noisy and unreliable, and it had too many parts to sit nicely in the temple. That's the main reason I moved to the ADS1292, which is built for biopotential sensing.

A lot of the build fought me on the first try. The thin temple body kept breaking until I thickened it and raised the infill, and the connector slot took a few iterations. Capacitive touch is the other thing I had to rethink, adding pads to this form factor made the temple thicker and harder to wear, so I left them out of this spiral and stuck with a simple button. The color display works but is probably overkill for what the glasses currently do, so I might drop it in a later spiral.

It gets evaluated on two fronts: the Fab Academy requirements, and my own test for whether it actually works.

On the Fab Academy side, the final project has to pull the required skills into one integrated object that is clearly mine and runs on its own. That means 2D and 3D design, both additive and subtractive fabrication, a microcontroller board I designed and produced myself with real input and output, embedded programming, and full system integration and packaging. NeuroAR hits each of those: the CAD frame and optics, the 3D-printed body with the laser-cut combiner and vinyl decals, my milled ADS1292 board with the electrodes as input and the display as output, the firmware, and the final packaged wearable.

I also tried to make rather than buy wherever it made sense. I made the board, the frame, the optics housing, and the combiner, and only bought the things that don't make sense to fabricate in a lab, like the XIAO, the ADS1292 chip, the lens, and the OLED. The BOM shows that split.

The documentation is part of the evaluation: the final project page, the bill of materials, all my original design files (CAD, board, and code), credit to the people and projects I built on, and the two presentation pieces, the slide and the one-minute video.

Then there's my own bar for whether it works. I want it judged as a worn, comfortable-enough object, not separate modules. It has to sit on my face, keep the display aligned, hold all the electronics, and survive being handled and power-cycled like a real consumer device. The XIAO should drive the display and pass the ADS1292 checks (registers reading back, DRDY behaving, clean sample frames over SPI), and the input should reliably catch at least strong forehead events like blinks and eye movement. Clean EEG and SNN-based detection stay as stretch goals.

The main thing I'm still not sure about is how far to push the EEG claim. Right now the honest version is wearable AR plus biopotential sensing, with clean EEG and SNN processing as the direction I explored, not a finished feature. There is quite a bit of noise and instability with the EEG signals, and using additional electrodes plus better isolation of my traces would probably go a long way. Since I can't mill thinner traces, I didn't explore the ADS1299 (a chip more targeted at EEG); ordering a ready PCB from a board house could be a good next spiral.

A software spiral I've already started is the AI and EEG side on the same hardware: a small SNN classifier in PSRAM to pull out brain-state classes like focus, attention, blink intent, and rest, used to drive the display with blink-to-click and alpha-gated menus, and feeding my assistant endpoint as extra context. With what I have now, without additional electrodes, I was able to run an SNN classifier that sort-of works.

A spiral on the current form factor is a hardware shrink: the electrodes and sensing move into the temples for shorter leads and less noise, the external powerbank gets replaced by an in-frame LiPo, and I design proper dedicated lenses. It would add smart-glasses features like a microphone, capacitive touch, speakers, and a camera, plus a dedicated enclosure for the electrodes.

Since that hardware spiral is complex, I want the system to ride on top of existing smart glasses like the Meta Ray-Bans through their SDK, so the useful features work on hardware people already wear.

What this project demonstrates is the ability to combine two increasingly adopted technologies in a single form factor. This integration could enable new kinds of human-computer interaction through interfacing between the mind and mixed reality, and it could act as a platform for people to build more advanced applications of monitoring brain activity with AR. I love thinking about bringing our brains into the equation of the devices we interact with daily and using data from them to guide how we use those devices.