Scribble.bio
A pre-product company, not a client build — a computer-vision and AI system that measures the structure of handwriting. The site is built as an instrument panel for that claim: a live, in-browser demo computes real geometry from the sample on screen, and every feature past that line is labelled exactly what it is.
Scribble.bio is Ishaan's own venture, built to prove a narrower and harder claim than most AI products attempt: that handwriting has measurable, reproducible structure, before anyone asks whether that structure means anything about the person. Hover a word in the live sample on the homepage and the page fits a baseline through the letters, measures slant weighted by how much ink each stroke contributes, groups ink into words, and reports spacing normalised against the writer's own letter height — numbers computed from the actual sample on screen, not staged. That live demo is stage one of a five-stage pipeline the product is engineered toward: ingest and quality control (resolution, lighting, blur and orientation scoring, reject-or-retake), deskew-and-normalise preprocessing, feature extraction across slant, spacing, size, pressure proxies, loops, baseline and rhythm, ensemble model scoring that maps those features to trait dimensions with confidence intervals against expert-labelled references, and an explainability layer that names the specific features behind every reading rather than asserting a conclusion. The rest of the roadmap — social discovery, compatibility matching, a vetted-graphologist marketplace, career coaching — is drawn on the site using the same handwriting-rendering engine as the hero, but every one of those sections is labelled precisely what it is: in development, coming later, future. Nothing on the site claims to be shipped until the measurement underneath it is real.
- 01A live, in-browser computer-vision demo — hover a word and see baseline angle, slant, stroke consistency, spacing and rhythm measured from the actual sample on the page, not staged numbers
- 02The full pipeline it's engineered toward: ingest & quality control, deskew/normalise preprocessing, feature extraction across slant, spacing, pressure proxies, loops, baseline and rhythm, ensemble scoring against expert-labelled references with confidence intervals, and an explainability layer naming the features behind every reading
- 03Every unbuilt feature labelled honestly — in development, coming later, future — instead of presented as shipped
- 04A confidence-scored reading model where 'limited evidence' is a valid, reported result, not a gap the copy papers over
- 05Concept sections for social discovery, compatibility matching and a vetted-graphologist marketplace, illustrated with the same handwriting engine rather than mockup screenshots
- 06A cinematic Three.js hero — a paper plane of handwriting that folds and responds as part of the page, not a static illustration
The site proves the one claim the rest of the roadmap depends on — that handwriting structure is measurable and reproducible — before any social, matching or expert-marketplace layer gets built on top of it. Everything past that foundation stays labelled as exactly what it is: in development, coming later, or future.