A seance-inspired installation powered by CoreXY robotics and generative AI
CoreXY gantry hidden beneath the board drives a neodymium magnet with ±0.5 mm repeatability across a 650 mm workspace.
Hands-free visitor flow pairs PyAudio capture, Whisper transcription, and a GPT persona to answer questions in under nine seconds.
A Raspberry Pi orchestrates audio and API calls while an Arduino running GRBL executes deterministic motion trajectories.
The AI Ouija Board fuses traditional planchette storytelling with modern robotics and language models. Visitors whisper a question, the system transcribes the audio, prompts a GPT persona, and then drives a concealed magnet to spell the reply letter by letter. The installation was built alongside Evan Park and David Vaughn during Fab Academy Machine Week.
The build emphasizes a self-contained loop: a Raspberry Pi 4B handles audio capture, Whisper inference, and OpenAI requests; an Arduino Uno running GRBL translates the resulting string into motion; and a CoreXY gantry hidden beneath the board delivers the theatrical reveal. One "commune" takes around 90 seconds.
We began with Jack Hollingsworth’s automated Ouija board and set out to remove the keyboard, laptop, and exposed UI. The final flow relies solely on audio prompts and ambient cues: visitors tap a capacitive panel, speak their question, watch the board “think,” and then follow the planchette as it answers in an intentionally ominous tone.
Persona prompts constrain GPT responses to stay in character and under twenty-five characters. Keeping replies concise limits traversal distance, protects the hardware from prolonged stalls, and ensures every session feels responsive.
The enclosure was parameterized in Fusion 360 to balance portability with enough interior volume for the gantry and electronics. We iterated with laser-cut cardboard mockups before milling the final birch panels on a ShopBot.
Beneath the surface, a CoreXY layout drives a magnet carriage. Both stepper axes contribute to most moves, so belt tension and driver tuning were critical: we tightened belts to 40–45 N using a tension gauge and tuned DRV8825 current limits to 0.9 A to avoid missed steps.
We verified backlash by jogging between fiducial marks and measuring the error with a dial indicator. The resulting ±0.5 mm repeatability is well within the letter-spacing tolerances of the engraved board.
A wooden planchette houses the magnet saddle. The saddle offsets the magnet to clear belt hardware, so the coordinate lookup table incorporates a 7.5 mm Y-offset. Interchangeable caps let us experiment with translucent pointers and engraved acrylic lenses.
Motion control is anchored by an Arduino Uno flashed with GRBL 1.1f on a CNC shield. Two NEMA 17 steppers and DRV8825 drivers, microstepping at 1/16, deliver 80 steps/mm. Dual limit switches square the axes, and the Pi communicates over USB serial at 115200 baud.
Each character on the board maps to a coordinate pair captured with a dial indicator. The helper below reads the lookup table, enforces two-decimal precision, and writes newline-terminated commands that GRBL expects.
const FEED_RATE = 2200;
function moveTo(letter: string) {
const [x, y] = alphabetMap[letter.toUpperCase()];
const command = `G1 X${x.toFixed(2)} Y${y.toFixed(2)} F${FEED_RATE}\n`;
serial.write(command);
}A Raspberry Pi 4B logs audio via PyAudio using 16-bit PCM at 22.05 kHz to minimize buffer overruns. Before recording starts, the script enumerates audio devices to pin the USB microphone index and verify supported sample rates.
Whisper provides dependable transcripts even in exhibition halls. The transcript feeds a GPT persona that is primed to stay in character and respect a strict character budget. We validated prompts with synthetic inputs to ensure tone and brevity before shipping to the show floor.
The Pi streams characters over serial, each resolving to a coordinate lookup and a buffered `G1` move. We inject a configurable 1.5-second dwell between letters so audiences can register each stop. Feed rates are capped at 2200 mm/min to prevent belt slip.
A watchdog timer pauses the sequence if no acknowledgment is received within the dwell window—a safeguard that proved useful when visitors brushed the planchette mid-run.
We staged the integration over a weekend: mechanical bring-up on Friday, electronics on Saturday, and software tuning on Sunday. Keeping each layer modular simplified debugging when we inevitably retraced steps.
Demo clip captured during FabAcademy Machine Week
Next steps include experimenting with multi-planchette storytelling, adding positional audio, and expanding the coordinate map to support numerology modes without sacrificing speed.
CAD, firmware, Python scripts, and documentation from FabAcademy Machine Week.