Nyx / Nadir · Own product · 2026 — ongoing
I designed and built a personal AI companion, then developed it into Nadir: a native app with distinct characters, persistent memory and a system for creating more.
Warmth with forward motion.
Curious, lively and gently direct. Her character makes room for a small next step without dismissing what came before.
Imagination with a point of view.
Calm, attentive and precise. He makes possibilities easier to compare, then helps turn them into a concrete choice.
Original Companion selection artwork. Eos and Morpheus are the two free starter characters in Nadir.
The design work sits between what a character looks like, how it responds, what it remembers and how much initiative it takes.
Nyx started as a companion for my own daily life. Nadir brings the same core to other people, with their own data and a choice of companions. I own the product direction, persona logic, interface, specifications and quality of the AI-assisted implementation.
I defined Eos and Morpheus through behaviour as well as appearance: how they disagree, respond to uncertainty and move a conversation forward. A shared set of expressions makes those differences visible while keeping each character recognisable.
Choosing a companion is the start of a relationship. The onboarding gives that decision its own space.
Full character cards introduce Eos and Morpheus together. Opening one reveals its voice, personality and visual world before the user commits.
Selection and introduction are separate states, with a clear way back. Both starter companions are included, and the choice can be changed later.
A blank chat field asks the user to invent the relationship. Presence gives them something to respond to.
A contextual greeting and three re-entry cards establish a starting point: an open thread, something pending or a conversation starter. Empty threads and tasks are named explicitly, so the first screen still offers a way in.
Deeper tasks live in focused workspaces with their own tools and permissions. That keeps the companion’s social presence coherent while still making it useful.
Closeness, memory and initiative only work if the companion holds its boundaries. Two decisions from everyday use shaped how I designed the product.
In one calendar answer, the model invented an appointment. A convincing response had become a reason to distrust the product.
I made the loaded calendar data the boundary: answers may only use events that are actually available. If an entry is missing, the correct response is “not found”. That made uncertainty a product state to design for.
Saving a bookmark should not depend on a model or network request succeeding. I separated capture from AI enrichment.
The link, text and images are stored locally first; summaries and tags follow in the background. If enrichment fails, the saved entry stays available. The interface makes that order explicit, so an optional service cannot take away the primary action.
The same principle shapes initiative: a stored API key is not blanket consent for background work, and an action with an outside effect needs its own scope and controls. The automation architecture below makes those decisions traceable.
Pose Studio turns one character master into a reviewed, consistent set of assets that the app can actually use.
ComfyUI did not cover the complete workflow I needed. I built a local Python tool around the repeated decisions: generate, compare, adjust, approve and assemble.
It supports Gemini image models and GPT Image via OpenRouter, plus manual import from tools such as Mage.Space. Each expression keeps its prompt, versions and approval state. A failed pose can be replaced without rebuilding the set.
Start from one visual master. Set the character, posture and expression prompts.
Build 42 expressions for each avatar and chibi set. Regenerate individual poses and explicitly approve the result.
Upscale the approved set, clean up the artwork in Figma and return the final transparent images.
Combine the artwork with persona, themes, stage and selection card. Validate the complete source before packaging.
The tool also assembles persona content, themes and the selection card. Incomplete or outdated image sets block export. The final package carries 42 chat expressions and 42 chibis per theme, giving the product a repeatable character-production process.
Nadir’s business model separates access to the product from the choice of character.
The app is free, with Eos and Morpheus included. Additional official companions are the planned paid offering. The product hypothesis is that people will pay for a character’s personality and visual world once the core experience has earned a place in their day.
That decision shapes the product: clear character presentation, a consistent content package and an explicit choice before switching. A new companion is imported as a signed file; installing it does not silently replace the active one.
Current stage: the package and import foundation is implemented. Paid distribution and willingness to pay remain to be validated.
React, FastAPI and Tauri form the shared product core. SQLite and a local vault hold the data; model providers are optional connections. Nyx runs privately on my infrastructure, while Nadir bundles its local backend in the macOS app.
The two editions share code, but not a data path. Edition profiles and bundle checks keep private Nyx content out of Nadir. Memory has an origin, a scope and a lifecycle, including deletion of derived indexes. Local-first means stored content and core functions stay useful without a provider; generative features still need a chosen local or external model.
These are design decisions as much as implementation details: what the companion recalls, when it speaks and how it handles uncertainty all shape the relationship.
The companion also sits beside focused workspaces. Document actions begin with a selected file; calendar prompts relate to events, free time and overdue tasks.
I use AI across the work: exploring images, building software, testing voices and connecting creative tools. When an existing workflow falls short, I build the missing part.
I use Nyx daily. Two people use Nadir in closed alpha.
Alpha feedback led to specific changes: persisted names became stable, the empty first start gained a clear way in, and optional provider errors stopped hiding content that had already loaded.
I specify behaviour and acceptance criteria, build in small slices with AI assistance, and review the result against the actual code and built app. Updates, permissions and preservation of local data are part of the product experience.
This stage gives me direct feedback on the experience. The next questions are broader adoption, repeat use and whether the additional companions earn their price.
Product status confirmed September 2026 · Independent project · Product design and AI-assisted implementation by Arne Feike