We make a handheld that types what you draw in the air. Since May, 450 paid volunteers have written for it across 24 studies; the largest set — 349 adults aged 18 to 78, 52,191 letters, with age, sex and handedness recorded — is where every accuracy figure here comes from. Tested the hard way, leave one person out and train on everyone else: 68% of a stranger's letters right, cold; with the context of a word, 85% of words exactly right — and half the people it has never met get 95%. Twenty-eight of them then wrote for it a second time, with a copy of the recogniser fine-tuned to each of them in between; that result, every number, is below.
What the wand is
Hold it like a TV remote, draw a letter, and a small neural network turns the movement into text. It connects as an ordinary Bluetooth keyboard, so a TV, a phone or a laptop needs nothing installed. No camera: recognition runs on a chip inside the device, reading a motion sensor that costs a few dollars.
Where the data came from
Three places. Paid volunteers on Prolific — a British online platform that university researchers use to recruit study participants: people sign up, are verified, choose the studies they want to take, and are paid for their time; their age, sex and handedness come from the profile they keep there. Since May, 450 of them have written for the wand across 24 studies. The largest set is 349 adults aged 18 to 78, who wrote every letter, digit and a few shapes three times each on their own phone, held like a remote — 52,191 letters, and every accuracy figure below comes from it. People playing our free games, Flags of the World among them: about 2,000 letters from 74 players in nine countries. And nine participants in an IRB-approved pilot study run by university researchers in the US, who wrote on our prototype hardware in spring 2026 — the first step of an independent evaluation; the complete, larger study is under way this autumn.
How well it reads a stranger
Tested the hard way — leave one person out, train on everyone else, see how it does on someone it has never met: 68% of letters right, cold; with the context of a word, 85% of words exactly right. Left-handers 68.5%, right-handers 68.2%. Women 68.0%, men 67.4%. No difference by age from 25 up; the under-25s are the exception at 55%, and a fifth of them drew letters a few centimetres across. The average hides a tail: half the people it has never met get 95% of their words exactly right, the better half 97%, and the worst tenth are at 65% or below — they are who the next section is for. The full analysis, with the method and the caveats, is the Air-Writing Atlas.
What your own writing buys
Twenty-eight of the volunteers wrote for it twice, on different days, the alphabet a few times over each time, with a copy of the recogniser fine-tuned to each of them in between. We said we would publish the result whichever way it went, and it went both ways: the twelve it struggled with went from 82% to 93.5% of words exactly right, all twelve up; the sixteen it already read well went from 95.7% to 96.5%, four of them slightly worse. Twenty-four of the 28 improved, and afterwards nobody was above 20% word error. Personalisation fixes the hands it struggles with and leaves the rest alone.
Try it
Open phone-demo.the-wand.ai/write on your phone and write a word in the air. Then tap “train it on your handwriting": draw the alphabet twice, about five minutes, and we email you a link to a copy of the recogniser trained on your hand. On the wand itself the same pipeline — we call it Attune, about fifteen minutes of writing — puts your model on the device over the air.
What more data buys
Less each time. With 25 people in the training set it read 48% of a newcomer’s letters; with 100, 61%; with 270, 69%. To reach 75% the fit says about 800 people; 80%, about 2,300. The 2,000 free letters from the games added nothing we could measure. We ran that twice.
The twenty-eight, one by one
Whole words exactly right, built from each person’s real recorded letters and run through the full pipeline — recogniser plus the correction model — on a standard English phrase set. Day 1 is cold: the recogniser had never seen that person. Day 2 is after a copy of it was fine-tuned on their first session only.
| # | cohort | words exactly right, day 1 (cold) | day 2, after it learned the hand | change |
|---|---|---|---|---|
| 1 | the twelve | 47.3% | 92.1% | +44.7 pp |
| 2 | the twelve | 58.8% | 87.2% | +28.4 pp |
| 3 | the twelve | 69.7% | 89.5% | +19.8 pp |
| 4 | the twelve | 79.5% | 92.4% | +12.9 pp |
| 5 | the twelve | 84.5% | 94.4% | +9.9 pp |
| 6 | the twelve | 84.8% | 95.3% | +10.5 pp |
| 7 | the twelve | 86.6% | 90.9% | +4.3 pp |
| 8 | the twelve | 87.2% | 92.7% | +5.5 pp |
| 9 | the twelve | 93.8% | 94.5% | +0.7 pp |
| 10 | the twelve | 95.2% | 96.7% | +1.4 pp |
| 11 | the twelve | 96.9% | 98.1% | +1.2 pp |
| 12 | the twelve | 98.3% | 98.4% | +0.1 pp |
| 13 | the sixteen | 90.1% | 93.8% | +3.6 pp |
| 14 | the sixteen | 90.5% | 89.6% | -0.9 pp |
| 15 | the sixteen | 90.6% | 94.6% | +3.9 pp |
| 16 | the sixteen | 93.4% | 93.2% | -0.2 pp |
| 17 | the sixteen | 93.7% | 92.5% | -1.3 pp |
| 18 | the sixteen | 93.8% | 95.5% | +1.6 pp |
| 19 | the sixteen | 96.1% | 98.6% | +2.6 pp |
| 20 | the sixteen | 96.4% | 98.6% | +2.2 pp |
| 21 | the sixteen | 96.9% | 97.3% | +0.4 pp |
| 22 | the sixteen | 97.5% | 98.5% | +1.0 pp |
| 23 | the sixteen | 98.1% | 98.6% | +0.5 pp |
| 24 | the sixteen | 98.3% | 97.7% | -0.6 pp |
| 25 | the sixteen | 98.4% | 99.0% | +0.7 pp |
| 26 | the sixteen | 98.5% | 98.6% | +0.2 pp |
| 27 | the sixteen | 98.8% | 98.9% | +0.0 pp |
| 28 | the sixteen | 99.1% | 99.3% | +0.1 pp |
The twelve came back on their own, the next day; the sixteen were invited back two months later. The twelve were a hard cohort — 18.1% word error cold against the population’s 14.9% — and the sixteen an easy one, which is most of why the two gains differ. Eighteen of the 28 invited back completed the second session; two of those are excluded because their phones supplied no rotation data in either session, so the analysed cohort is twelve plus sixteen.
Nine things to take away
Three numbers. 450 — the paid volunteers who have written in the air for it since May. 85% — the words a stranger’s writing comes out exactly right, first time; 97% for the better half of people. 82% → 94% — after five minutes of your own writing, for the people it found hardest.
Three things about us. No camera — a handheld that types what you draw in the air; a chip inside, an ordinary Bluetooth keyboard. Five patent families — filed; three co-founders: audio and edge AI, commercial strategy, an IP attorney. CES in January — production boards in build; independently evaluated by university researchers in the US.
Three for a dinner party (from the Atlas). Lefties go clockwise — a third of left-handers’ circles run clockwise; right-handers’, 15%. 27 cm and a metre — an air-letter is about 27 cm tall, takes under two seconds, and the hand travels nearly a metre. Postage stamp — one in five under-25s write letters that small; it reads the over-60s better than the 20-somethings.
What’s next
This autumn the same university researchers run a complete, larger study with the latest hardware — how quickly people learn it, how the prompt shapes the stroke, where the button should sit. Peer-reviewed publications are in preparation and we plan to open-source the data. The first production-design boards are fabricated and we plan to show working hardware at CES in January. Nothing is on sale — no price, no ship date. The plan for 2027 is a crowdfunded first run of the device alongside licensing the recogniser to other device makers.
My parents’ handwriting has been their own for as long as I’ve known it. The machine should learn them, not the other way round.
— John
Method, briefly
The 450 are the unique paid participants approved across our 24 air-writing studies on Prolific (2026-05-07 to date), from the studies' own demographic exports; a further 44 people took part in a separate launch-animation vote and are not counted. Cold accuracy is leave-one-subject-out: the model that reads a person was trained on everyone except that person. The 68% figure is over 48 prompted glyph classes; the 85% of words is the 37-class letter model plus the WandSpell context decoder over standard English test phrases assembled from each participant's real recorded letters, across 317 held-out people, and the 95% / 97% / 65% figures are the median, the better-half mean and the 90th-percentile point of that same distribution. The curve is a fixed held-out panel of 40 people never in any training set, three random draws of the training set per size, cold start, 48 glyphs; the dashed part is a power-law fit, not a measurement. The games' null: adding the arcade letters to the training set moved held-out accuracy by −0.34 pp (p 0.20, 152 people) and then −0.05 pp (p 0.63, 108 people). The 28: a frozen personalisation recipe with no tuning on the evaluated people; the twelve's cold figure reproduces exactly (18.10% word error) against the earlier publication; the recipe's rehearsal pool excludes every evaluated subject. Never merged: cold-start and personalised figures, and population and single-person figures, are reported apart. Sources: experiments/2026-08-17-personalization-n30/RESULTS.md, experiments/2026-08-11-letter-personalized-wer/, experiments/2026-06-13-loso-full-runs1-5/, experiments/2026-08-15-air-writing-atlas/REPORT.md, wand-study-data/_demographics/.