ClipSift v0.1.0 Alpha

ClipSift v0.1.0 Alpha

ClipSift's first Alpha pre-release is a privacy-conscious Windows 10/11 64-bit CCTV review assistant powered by the vision-capable google/gemma-3-4b-it model.

• Download: ClipSift-0.1.0-win64.zip — 2,876,584,258 bytes (2.68 GiB).
• SHA-256: ce3380b629086c00930928d0038e73f805d1a57f93794ebaafd95d6a1e6bd525 (checksum file).

Highlights

• Desktop GUI: Select input/output folders, device, sampling strategy and frame limit; follow scan progress; cancel safely; filter results; preview evidence; and open evidence or source video.
• CLI: Scan folders, inspect system readiness, and run controlled single-image or single-video tests using the same core workflow as the GUI.
• Local folder scanning: Read .mp4, .avi, .mov, and .mkv CCTV clips without modifying originals. Hybrid, uniform, and motion-prioritised sampling cover the readable timeline; motion selects candidate frames but is not treated as evidence of a person.
• Clear classifications: Deterministic application policy converts Gemma observations into Person Detected, Needs Review, or No Person Detected. Uncertain, malformed, and low-confidence assessments require review.
• Evidence and reports: Save the frame that triggered a person/review result, copy flagged/review clips, and produce report.csv and benchmark.json in the selected output folder.
• System Check: Inspect Python/runtime components, CUDA/GPU/VRAM, BitsAndBytes, Hugging Face authentication, and local model-cache signals without loading or downloading Gemma.

Tested environments

• NVIDIA RTX 3060 Laptop GPU: controlled Gemma image inference was exercised with the 4-bit NF4 safe preset.
• Google Cloud NVIDIA T4 environment: the ClipSift workflow was exercised as part of pre-release development.

These are compatibility observations, not a broad hardware certification or an accuracy/performance benchmark. Real inference with the final packaged executable remains a release-checklist item.

Packaging and model delivery

The Windows release is packaged as an unsigned, portable PyInstaller onedir/windowed application and distributed as a ZIP rather than an installer. The directory contains ClipSift and its Python/GUI/computer-vision/ML runtime, so it is large and must remain intact after extraction.

Gemma is not included in the ZIP, bundled, or redistributed. Users must have a Hugging Face account, accept the terms for google/gemma-3-4b-it, authenticate with hf auth login, and allow Hugging Face to download the model separately into its normal per-user cache. Credentials are not included. An NVIDIA GPU is recommended, and approximately 20 GB of free disk space is recommended for the application, model cache, and results.

Known limitations

• This is Alpha software and may contain defects or change incompatibly.
• Automated classifications are review aids, not proof. Low light, obstruction, weather, reflections, motion blur and compression can cause errors.
• There is no facial recognition, identity/intent inference, live-camera support, real-time alerting, installer, or automatic update mechanism.
• Practical inference currently expects a compatible NVIDIA CUDA setup; CPU inference may be extremely slow and memory-heavy.
• The gated model requires internet access for initial download, sufficient disk space, accepted Gemma terms, and valid Hugging Face authentication.
• The Windows executable is not code-signed and may trigger Microsoft Defender SmartScreen. Verify the published SHA-256 and obtain downloads only from the documented release location.
• System Check is diagnostic and offline; passing it does not establish successful real inference.