In the figments of your imagination, a world appears — picturesque scenes that flow like a river. And every second, it fades. The pen is too slow for the palace in your mind.
Our essays have autocomplete.
Shouldn’t our imagination?
MaoBi is autocomplete for drawing — digital paper with magic in it. Sketch a rough page, press Enter, and real, editable components bloom in like wet ink, exactly where you drew them. Press Frame, and the sketch becomes a working website.
The problem
Fast but fading, or faithful but slow
Drawing is the least restrictive way to get an idea out of your head — no syntax, no menus, no blank-page paralysis. But every way of capturing it loses something. Engineers, students, artists, designers — everyone who thinks faster than they can draw makes this trade.
Pen & paper
Fast and free — but the result is dead ink. Not structured, not editable, not shareable.
Design tools
Real structure — but menus, component browsers, and precision dragging interrupt thinking.
Text → image AI
A pretty picture — uneditable, detached from what you drew, no spatial control.
MaoBi
The speed of sketching, the structure of a design tool. Fast and faithful — for the first time.
How you drive it
Plain shapes stay shapes. A single letter adds function.
Every enclosed shape gets crisp and keeps the color you shaded it. Write one lone letter inside a box and it becomes a working component — words and colors around it become labels, fills, and gradients. No surprises, ever.
Button“Login” labels it
Navbarsnaps full-width
Formworking inputs
Image framedrop a photo in
Videoa real player
Placeholderdecide later
Under the hood
Two models, strict separation of powers
No model is ever trusted with geometry. Each layer does the one thing it’s good at — and every hand-off is validated, so a hallucination becomes nothing, never a broken page.
Your ink
Canvas
Screenshot + per-stroke manifest, colors included.
The eyes
Gemini vision
Describes only: kind, glyph, text, colors. Never places, never decides.
The ruler
Normalizer
Pure code. Geometry from your strokes; containment — what’s inside what.
The brain
FreeSolo 2B
Our fine-tuned model decides: op + params. Zero coordinates in its output.
The law
Validators
Schema, geometric, domain — fail closed. Junk output means nothing happens.
The hand
Renderer
Deterministic seeded templates. The model can’t draw an ugly button.
Zero coordinates, by design. The model never says where — geometry always derives from your actual ink, so nothing can ever be misplaced. And when it isn’t confident, it abstains: your ink stays ink.
The receipts
We trained the brain ourselves — and it beats Gemini
The decision-maker is Qwen3.5-2B, LoRA-fine-tuned on FreeSolo against our own synthetic sketch dataset. Scores below are the untouched held-out test split, our model vs the Gemini Flash baseline.
96.7%
Op accuracy
vs 75.0% Gemini Flash baseline
+33
Detail routing (pts)
word-in-box → label, colors → fill: 90–93.5% vs 58.7%
100%
Night-sky-from-rect
vs 25% baseline
0.97
Abstention F1
vs 0.67 — it knows when to stay quiet
<$0.25
Total training spend
~10 runs · 640 examples × 4 epochs · LoRA SFT
0
Broken pages possible
0–1.7% hallucination, all caught by fail-closed validators
Full eval ledger, test bank, and training configs live in the repo — every number is reproducible with scripts/eval-harness.ts.