screen lens dgx.stub

---

id: screen-lens-dgx name: ScreenLens-DGX repo_path: ~/Documents/sw30labs/repos/screen-lens-dgx remote_url: https://github.com/sw30labs/screen-lens-dgx.git primary_language: Python framework: LangGraph + vLLM + OpenCLIP + ChromaDB category: local-inference-dgx stacks: [dgx, langgraph, rag] ingested: 2026-07-22 last_commit_date: 2026-07-20

ScreenLens-DGX

Purpose

Local video scene intelligence for NVIDIA DGX Spark (GB10 / Linux aarch64) — the DGX-only fork of the dual-platform screen-lens. Extracts meaningful frames from screen recordings, captions or transcribes them with a local multimodal vLLM model, embeds them with OpenCLIP on CUDA, indexes them in ChromaDB, and reconstructs visible code, documents, PDFs, and GUI walkthroughs. Defaults are always vLLM + CUDA + concurrency 2; there is no macOS/oMLX product path.

Architecture

Same LangGraph StateGraph pipeline as the Apple sibling (hybrid keyframe detection SSIM + pHash + HSV → caption → OpenCLIP embed → ChromaDB index → classify/plan/reconstruct/QA with reflection retries), with the captioner pointed at vLLM on :8000 serving Qwen3.6-27B-FP8. Ships compose.dgx-spark.yaml for a Docker path plus setup_and_run_dgx.sh; TUI screenshots in assets. Recent work hardened OCR progress handling (stopped progress-bar loops, raised smoke-test timeout).

Dependencies

langgraph, langchain (+ollama, +community), chromadb, open-clip-torch, opencv-python-headless, scikit-image, Pillow, numpy, torch, tqdm, typer, rich, pydantic; extras dev / tui / all. Package screenlens-dgx 0.2.0; 7 test files.

CI/CD

No CI workflows in-repo. Docker compose for DGX Spark; Apache-2.0, v0.2.0.

Entities and topics

• screenlens — Apple Silicon / oMLX sibling (renamed screen-lens-mlx on GitHub)
• qwenbench-mlx — Qwen model-family benchmarking on the Apple side
• Tags: [python, langgraph, cuda, vllm, local-inference, chromadb, video, ocr, qwen]