Creativity, Craft, Computation

constantin.studio is a design studio led by Constantin Chopin — founder of constantin.world, faculty at ArtCenter College of Design.

We work with ambitious teams building AI-native products across consumer apps: Replika, enterprise intelligence layers: Parable, Pendulum, and cultural technologies: Lightnote.

The studio operates at the intersection of three crafts: the stories brands tell, the products people use, and the systems that run beneath them both. We move fast across mediums, iterate often, and sweat over the details.

Replika

Design engineeringDesign systemsConsumer mobileMemory interfaces2026

Role: Contract design engineer
Team: Flat 20 · 4 design, 10 eng
Shipped: Memory V2, onboarding
Reach: 42M users

Replika is the AI companion, with 42 million users worldwide. In 2026, mid-way through a full rebrand under a new CEO and an evolved vision, they brought us in to lead design engineering on a flat team of twenty, working across iOS, web, brand, and the design system.

The work was to define the flagship features of the product's next era, and we shipped directly with the lead iOS developer, carrying each feature from direction to spec to build.

The harder question was what that next era should be, decided in the middle of a live rebrand with the brand and its vision still in motion. Rather than start from features, we followed how people actually used the product back to a single reframe: memory is the product. That call set the direction for everything after it.

Memory V2 grew into Portraits, Life Arcs, and People Cards, all built on structured memory drawn from conversation. Alongside it we delivered the 2026 design system grown from the brand book, a 40-screen onboarding recast as a six-act narrative, and the dual-currency economy through which users shape the relationship.

Parable

Design systemsDesign engineeringAI-operable componentsEnterprise UX2025 — 2026

Role: Sole product designer
Method: Figma → code
The cut: ~53% of tokens
Outcome: A named space of valid output

Parable is the intelligence layer for enterprise operations. It reads the whole company and surfaces where the work, the waste, and the automation hide.

In September 2025 the founders brought us in as their sole product designer. They had engineers on staff but no product director, and only a loose direction to give shape to, so we designed an end-to-end journey for the ideal customer, prototyped first in Figma and then in code beside the front-end engineer.

The real problem turned out to be drift. With no system beneath the visual language, every sprint widened the gap between what shipped and what cohered, and for an AI to operate the product, taste had to become rules. So the audit cut roughly 53% of the token set, tightening it until a model could compose from a named space of valid output.

The hardest call was how to carry density. Our first attempt put a density prop on every component, which worked in isolation but fell apart in composition, where a compact table inside a comfortable card had two sources of truth. We moved density out of the component tree and onto the DOM, resolved through a single attribute cascade, because in a system an AI recomposes, composability wins. We delivered that foundation, and the remaining budget moved to engineering to build from it.

Lightnote

Creative toolingCanvas interfacesRapid prototypingFine-tuning workflows2024

Role: Product designer
Context: Post-pivot tooling
Method: Running prototype
Outcome: Open source, MIT

Lightnote is a creative AI toolkit that lived in the terminal and was given a canvas. Its sets, processes, and runs are arranged as rooms rather than commands, for assembling image sets, composing workflows, and training LoRAs.

It began weeks after the Pendulum pivot, when the lab needed its own fine-tuning tooling. We built a running prototype against the live engine, so the sketch became code and stayed there.

The brief arrived as a UI problem, put panels over the model calls, but the real problem was grammar. The CLI already spoke a complete language of sets, processes, and runs, and an interface that invented its own vocabulary would have forked the product in two. An early direction gave each model family its own bespoke panel, which looked richer on screen but broke that grammar into three dialects.

We rebuilt around one node form with flags as fields, letting models differ only in their parameters, and kept parity enforced by the API rather than by discipline, so a trigger word lands in the same call as its --trigger flag and nothing drifts. We delivered the canvas the lab itself worked in. Operations later paused at the investors' call, and in 2026 Page Nineteen released it publicly, open source under MIT.