Machine learning · 2026
Aphelion
Make the first draft. Keep control of every detail.
Aphelion brings AI assistance into an editor you can keep working in. Generate a starting point, revise an individual page and adjust the result with familiar layout tools. The document remains structured and editable, with local models and a saved history behind each change.
- Status
- Active development
- Domain
- Native design software · Local AI
- Role
- Product, editor architecture, rendering and local AI
- Year
- 2026
Explore the workflow with illustrative content.
Begin with a structure
A brief and a set of content blocks give the page its starting structure. Choose a layout below to explore another arrangement.
Compose an editable page
The middle layer arranges the content into a page. Switching layouts moves the blocks rather than replacing the page with a flat image.
Refine the selected element
The foreground shows the selected object and its place in the page. Separate the layers to see content, layout and edit controls as distinct parts.
Illustrative content · no live data
An idea becomes a document you can keep editing
Stack
- C++20
- Skia
- Dear ImGui
- SQLite
- llama.cpp
- Qwen
- libgit2
A native design editor for reports and presentations, combining editable pages, local AI generation and a document history that keeps every change recoverable.
From a web editor to a native application
Earlier work on InterfaceGenie, later called Lunar, explored a collaborative web design tool.
Aphelion takes that report-and-presentation direction into a separate native C++ application, with a new document format and local models at its centre.
Keep the result editable
A document is a tree of pages, text, shapes, images, tables and charts.
Both manual changes and AI revisions use the same change operations, so the history and undo controls continue to work after a generated edit.
Generate a page, then improve it
The current pipeline develops a plan and builds pages individually.
Layout examples give it concrete visual references, while structural checks catch problems before a page is accepted. A conversational edit can then target a particular object or page instead of recreating the whole document.
One rendering path
Skia draws the document on screen and supplies the PDF rendering path.
The same underlying layout supports export checks, helping catch differences between what is edited and what is delivered.
A working editor, still in development
The project now includes a native canvas, structured document storage, local generation and revision, import and export work, and automated editor checks.
Output quality and the visual review loop remain active development areas.
Pipeline
Brief to an editable document
Stage Can reject
-
01
Brief
Document plan
-
02
Document
Editable pages and objects
-
03
Local generation
Page drafts
Structured output from local models
-
04
Review
Checked and revised pages
-
05
Export
Rendered document
Shared Skia rendering path
Behind the work
The skills
that came together.
Editor
- Document and page modelling
- Undo, redo and revision history
- Rendering, layout and export
AI
- Local inference orchestration
- Page-scoped generation and revision
- Structured output validation
- Layout retrieval and exemplar selection
Project highlights
-
Editable from the start
Pages and objects remain accessible after generation
-
Local AI
Generation and revision run through local models
-
Recoverable changes
Manual and AI edits share document history
-
Native document format
Content, assets and history stored together