Kinetix

Unified inference monorepo for the NeoZedAtlas runtime stack.

Current layout:

• Build new shared runtime layers under engine/.
• Move modality-specific execution code into adapters/.
• Keep public package and orchestration entrypoints under sdk/.
• Keep executable entrypoints under apps/cli and apps/services.
• Keep local model artifacts under models/: models/text, models/ocr, models/vision
• Keep local datasets and evaluation assets under datasets/: datasets/vision
• Converge toward one scheduler and one reusable inference core for vision, OCR, text, video, TTS, and future AI runtimes.

Current directory intent:

• sdk/: package surface and shared execution/session entrypoints
• engine/: reusable runtime internals grouped by domain (core, artifacts, runtime, scheduler)
• engine/runtime/vision/: layered by concern (io, analysis, memory, modules, nn)
• engine/runtime/text/: keep generic decoder infrastructure at the root, and keep model-family-specific adaptation and optimization under families/<model>/
• adapters/: modality adapters (text, vision, ocr)
• apps/: executable surfaces only