aanikaatluri/perceptron_incident_reporting

Workplace Safety Video Analytics

Enterprise-style safety analytics over short security clips, powered by Perceptron Mk1.

Flow A — Safety Incident Review

Upload a security clip → structured JSON report with timestamped events, severity, visual evidence, and recommended actions. Output is constrained to a Pydantic SafetyReport schema via pydantic_format(). A fillable Workplace Incident Report PDF is generated from the same analysis (JSON output is unchanged).

Hugging Face Spaces setup

1. 1.Create a new Gradio Space on huggingface.co/new-space.
2. 2.Push this repository (app.py, analyze.py, models.py, requirements.txt, README.md).
3. 3.In Settings → Repository secrets, add PERCEPTRON_API_KEY.
4. 4.Open the App tab after the build finishes.

Local development

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env   # add your PERCEPTRON_API_KEY
python app.py          # Gradio UI at http://127.0.0.1:7860
python safety_check.py /path/to/video.mp4   # Flow A CLI
python compress_video.py /path/to/clip.mov    # convert + compress for upload

Limits

• —Video must end up as MP4 under ~15 MB (API request body cap is 20 MB). MOV and other formats are auto-converted when you click Analyze (requires ffmpeg; included on HF Spaces via apt.txt).
• —Perceptron meaningfully samples the first ~2 minutes of each clip; oversized uploads are trimmed automatically when compressing.
• —CLI prep: python compress_video.py your_clip.mov (requires ffmpeg).

Langfuse observability

Set LANGFUSE_PUBLIC_KEY, LANGFUSE_SECRET_KEY, and optionally LANGFUSE_BASE_URL in .env or Space secrets.

Each analysis run traces:

• —Flow span (flow-a-incident-review) — explicit trace I/O (video metadata only, no raw bytes), structured JSON output, event counts
• —Generation span (perceptron-mk1) — nested model call with prompt and clip/error summaries
• —Session grouping — Gradio session_hash propagated via propagate_attributes
• —User feedback scores — thumbs up/down (user-rating boolean score) with optional comments

Use Langfuse to filter low-rated traces, build annotation queues, and export datasets for regression testing.