Anastasiia Shulga.
  • Work
  • About
  • Contact
design@anastasiiashulga.com© 2026 Anastasiia Shulga
All Work
AI ProductSelf-initiated

Marrow

A brand and product system for a research lab building calm, honest AI tools. Variable identity, full design system, and the core product UI.

Role

Brand, Product, Motion

Year

2026

Scope

Identity · UI · Motion · System

marrow

Research · Tools · Trust

The 30-second version

Marrow is a brand and product system for a research lab building calm, honest AI tools. The work covers a variable wordmark, a complete identity system, a marketing site, and the core product UI. It’s a counterpoint to the current aesthetic of AI: less arena lighting, more reading room.

Problem

AI brands look hostile. The products feel worse.

Most lab identities today read as crypto: heavy gradients, aggressive geometry, a tone that says “we are powerful and you are not.” Then you open the product and find a chat box on top of a void. There's no sense of trust, no visible reasoning, no place to think.

Marrow exists to answer a different question. What would an AI brand look like if it were built for serious users, like researchers, who want a tool that respects their attention?

Naming

The substance, not the surface

Marrow — the substance inside the bone. The marrow of the matter. A research tool isn’t about the chrome; it’s about what it lets you pull out of the source. Naming the lab after the essence sets the tone for everything else.

Approach

A library, not a launchpad. Build the brand around restraint, and let the product carry the proof.

Wordmark

A mark that breathes

One glyph carries the variable axis — the “o”, the nucleus. It shifts in weight on a 3.5s cycle, subliminal proof the system is alive. Lowercase. No trailing punctuation. The wordmark earns attention by sitting still, not by signaling.

marrow
wght 500

Drag the slider to set the wght axis manually, or let it breathe.

Tokens

A small palette, used with discipline

surface.base

#0E0F12

surface.panel

#15171C

ink.primary

#E8DDD0

sage.primary

#9DB5A8

sage.deep

#5C8C7E

ink.secondary

#9A9590

Typography

Serif for thinking, mono for evidence

Tools that show their work.

Display · Syne 500 · -2% tracking

Body copy stays out of the way. The eye should land on conclusions, not on the styling around them.

Body · Inter · 17/29

source.id = “pattyn-2023” · weight = 0.94

Evidence · SF Mono · 14

Product

The interface as proof

Three patterns do most of the work. A breathing status that shows the system thinking. An explicit source rail under every claim. A reading-first layout, with the response weighted heavier than the input.

marrow.research
marrow

Threads

Synthesis on climate modeling
Re-read of Whitehead
Bayesian priors draft
Architecture notes
A

What does the recent literature say about ice sheet sensitivity to ocean heat content above 2°C of forcing?

Synthesizing · 14 sources
2.4s

Recent work (Pattyn 2023, Slater et al. 2024) finds a non-linear response in West Antarctic ice loss when subsurface ocean temperatures exceed roughly +1.8°C above pre-industrial.

The mechanism is melt-water induced ice shelf weakening, which removes back-stress on grounded ice.

Pattyn 2023Slater 2024IPCC AR6 Ch.9+11
Ask a follow-up…⌘ ↵

Marrow Workbench · synthesis view

Motion

Three behaviors, applied consistently

Breathing

3.5s · sine

Idle. Used on the mark and on the synthesizing indicator. Proof the system is alive.

Settling

180ms · ease-out

When content arrives. Never bounces. Direction over decoration.

Reveal

60ms · stagger

Source tags and cards enter sequenced. Reading order made physical.

Applied

The mark across surfaces

m

Favicon · initial mark

m

App icon · accent ground

marrow

Header · in-product

What I'd do next

The hard part is the empty state

A blank prompt is the moment most AI tools fail. They show capability dumps, sample queries, or a void. The next phase of Marrow is a designed empty state: a small selection of the user's recent threads, a single suggested re-entry point, and nothing else.

I'd also push the explainability layer further. Right now the source rail is honest but flat. Hover states that surface the specific passage being cited would close the loop between claim and evidence.

← PreviousOctavoEditorial ProductNext →Identity in MotionIdentity