What problem does it solve? GDM cargo-train inspections for DeAcero arrive as Gmail emails with Excel label sheets and multi-page PDF photo evidence, and manually renaming photos, reading car IDs, and judging tampering (MANIPULADA) across thousands of cars is slow and error-prone. This Skill provides the tooling and hard-won domain knowledge to automate that pipeline end to end. ## Core Features & Use Cases - Attachment extraction and parsing: Pull PDF/Excel attachments from Gmail via the REST API, extract per-car cargo photos with PyMuPDF, and join them to ESTATUS labels by normalized PLACA. - Two-regime theft detection: Route photos by covered vs uncovered regime and apply per-regime ResNet models (uncovered ~96% recall, covered as a recall-first flagging aid), selected by Average Precision. - Hybrid PLACA OCR: Tesseract-first with a vision-LLM fallback and double-read consistency check, reaching 92% match with 0% wrong-car error. - Report generation and human review: A FastAPI reviewer app that drafts Spanish Excel/PDF reports matching GDM's exact templates, with human approval before emailing. - Use Case: Given a new inspection email, extract the cargo photos, score each car for tampering, surface upstream checkpoint history, and produce a reviewer-approved Excel and PDF report for the client. ## Quick Start Ask the agent to pull the latest GDM inspection report from Gmail, extract the cargo photos, and run the theft-detection models to draft a review-ready report.
