Can I vibecode Shade?

Act as an expert AI engineer and full-stack developer. Build a comprehensive local media asset management system and semantic search engine designed to replicate and extend the core functionalities of Shade, tailored for video editors, photographers, and creative teams.

System Architecture & Requirements:
- Develop a Python-based CLI application coupled with a sleek, responsive FastAPI web interface served on localhost.
- Utilize SQLite and sqlite-vec to robustly handle both metadata and high-dimensional vector embeddings within a unified, single-file database (`~/FootageIndex/index.db`).
- Implement an `index <directory>` command that recursively scans the filesystem. For video files, employ FFmpeg to extract keyframes at 5-second intervals as well as dynamically on every detected scene change using the `scenedetect` or FFmpeg scene filter.
- Leverage the `open_clip` model (using ViT-B/32 or a more performant modern alternative) to generate vector embeddings for every extracted frame, stored alongside precise timestamps. Treat standalone images (stills, photos) identically to single video frames.
- Process all audio streams through a local Whisper implementation (e.g., `whisper.cpp` or `faster-whisper`). Generate highly accurate, timestamped transcripts. Chunk transcripts into overlapping 30-second temporal windows and embed them using a robust `sentence-transformers` model to enable deep semantic text search.

Search Capabilities & UI/UX:
- Build a natural language search interface allowing users to input complex descriptive queries (e.g., "a person running through a forest at sunset" or "drone shot of city traffic").
- The backend must seamlessly fuse image and transcript vectors, returning cross-modal, ranked search results presented as high-resolution thumbnails, file names, and exact timecodes.
- Ensure clicking a result instantly launches a robust local media player directly at that specific frame.

Indexing & Reliability:
- The indexing process must be strictly incremental, resumable, and uniquely keyed by file path, last modified time (`mtime`), and file size.
- Ensure the tool gracefully handles large libraries, printing continuous progress updates so long-running (overnight) tasks can be reliably audited.

Privacy & Local-First Constraints:
- The system must operate 100% locally. Include all necessary model weights. No cloud dependencies, no API keys, no telemetry, and zero data exfiltration.
- Treat all indexed files as strictly read-only; never modify, move, or rename source footage.
- Exclude authentication, server deployments, user permissions, and remote sharing features—this is a strict single-user application.

Documentation:
- Provide a robust README detailing setup, ffmpeg/whisper installation, initial model download steps, and an accurate estimation of indexing throughput (hours per terabyte) on both CPU and GPU hardware.