Aether AI Screenshot Studio

Autonomous, private, self-healing screenshot intelligence.

Designed & built by

NIKHIL CHARY SRIRAMOJU

GitHub · LinkedIn · sriramojunikhil66@gmail.com · +91 6300556301

Quick start

# 1. Zero-install demo (100% in-browser): open demo/index.html
# 2. Full stack
cp .env.example .env && docker compose up --build      # web :3000 · api :8000 · collab :8001
# 3. Backend only, no Docker
cd backend && pip install -r requirements.txt && uvicorn app.main:app --reload
# 4. Tests (numpy + opencv + pillow + pyjwt only)
cd backend && python -m tests.test_core

Enable the ML tier with docker compose build --build-arg WITH_ML=1 (or pip install -r requirements-ml.txt) and set AETHER_WEIGHTS once you have trained weights.

Live editor (index.html)

Open index.html (or the GitHub Pages link). It edits PDF, PNG, JPG, WebP, GIF, BMP, AVIF, SVG, HEIC, TIFF and saved project files, entirely in the browser, and exports PNG, JPG, WebP, PDF, SVG or a project file. Highlights: text replacement that copies the original colour, weight, size, edge softness and (for PDFs, exactly; for images, by OCR + font fingerprinting) the font; smart erase that rebuilds gradients and film grain; on-device OCR text editing; private-info detection and redaction; translation to 30 languages (Chrome's on-device Translator when present); WCAG AAA audit and fixes; dark-mode variant; colour-blindness preview; crop, extend, multi-page PDFs, voice/typed commands, code export. New: real vector-editable PDF export (strip the original glyphs from the PDF's own content stream, redraw the replacement in a matched standard font, keep every path/image untouched — proven against real pdf-lib and poppler, not just rasterised) for pages you only text-edited; a per-page eligibility panel in the Export tab shows which pages qualify and why; batch "translate all pages" and "redact all pages"; keyboard shortcuts (1-7 tools, [ ] pages, Ctrl+E export). Tests: node tests-web/test_core.js && node tests-web/flow_test.js && node tests-web/cine_test.js && node tests-web/pdf_export_test.js (the last one needs pdf-lib installed locally and qpdf/pdftotext on PATH).

What it does

Upload a screenshot -> detect layers (text, buttons, cards, images, icons) -> extract palettes/gradients/shadows -> match fonts & tokens via RAG -> edit (smart erase, outpaint, text replace, translate to 50+ languages, dark mode) with a self-healing render loop -> audit accessibility (WCAG AAA, colour-blindness sim) -> export React/Tailwind, HTML/CSS, Vue, SwiftUI, Flutter, Compose, Figma JSON -> collaborate live (Yjs CRDT, cursors, branches). Natural-language and voice commands drive all of it; every step lands in a live audit log.

See docs/ARCHITECTURE.md for the sequence diagram, feature map and API.

Status — read this

Verified by tests here	Written but not executed in the build environment	Needs your data/compute
contrast/AAA maths + auto-fix, i18n, classical layer detection, all 7 exporters, outpaint, erase, text replace, self-heal convergence, NL parser, dark mode	FastAPI app, Next.js UI, Yjs collab server, Celery task, ONNX loader, Docker/CI	CNN weights & font-classifier training, OCR wiring, 1M-font corpus, diffusion inpainting model, translation provider (glossary covers 5 languages x 7 UI words)

Repo map

backend/app — main.py API · engine.py · cv/ · rag/ · agents/graph.py · services/ · collab/ · auth.py · branding.py frontend — Next.js 14 App Router studio UI · demo/ — offline demo (GitHub Pages ready)

© NIKHIL CHARY SRIRAMOJU