Vision AI Engineer — Build 99% Accurate Playing Card Detection

I have a working Python/Flask web app deployed on Railway that detects playing cards from real-world smartphone photos. The app accepts an uploaded image, sends it to Google Gemini Flash Vision API, and returns the rank and suit of every detected card as structured JSON. The app already works end-to-end. The main challenge is accuracy. I need an experienced Gemini Flash / vision AI / prompt engineering specialist to help improve the system to 99%+ F1 accuracy on labeled real-world phone photos using Gemini Flash or a similarly low-cost vision model/API. The goal is to improve accuracy through smarter use of Gemini Flash or similar models, supported by better prompting, image preprocessing, structured validation, region-specific rechecking, and multi-step inference logic. The current system struggles with issues such as suit color confusion, diamond vs heart mistakes, rotated corner reading errors, overlapping cards, duplicate cards from multiple decks, and adjacent face cards being merged. I need someone who can improve the Gemini Flash workflow and help reduce these errors while keeping the solution cost-efficient and production-friendly. The final solution should remain suitable for repeated real-world use, with low marginal inference cost and clear documentation of what techniques were tested and what accuracy improvement each technique produced. Deliverables - Improve the existing Gemini Flash-based playing card detection workflow inside the current Python/Flask app. - Improve detection accuracy toward 99%+ F1 on the labeled real-world smartphone image set. - Create a stronger prompt structure that can detect all visible playing cards, identify rank and suit accurately, and return clean structured JSON output. - Improve handling of known failure cases, including black/red suit confusion, diamond vs heart confusion, 6 vs 9 rotated corner mistakes, overlapping cards, duplicate cards from multiple decks, and merged adjacent face cards. - Add useful image preprocessing improvements where needed, including lighting normalization, color correction, glare/reflection reduction, sharpening small card-corner details, and cropping or zooming image regions before Gemini Flash inference. - Build a multi-step Gemini Flash workflow if needed, including a full-image detection pass, region-specific recheck pass, duplicate validation pass, uncertainty correction pass, and final structured JSON validation. - Provide updated evaluation results showing F1 score, precision, recall, per-image performance breakdown, and before/after comparison. - Document all important implementation details, including techniques tested, measured F1 impact of each technique, final prompt/workflow used, new dependencies if any, environment variables if any, Railway deployment steps, and estimated cost per detection. - Provide a short operational guide explaining how to maintain the solution, how to run evaluations, and how users should take photos for best detection accuracy. - Walk me through all changes live so I understand the implementation and can maintain it later.