🌱 Seed
Automate the brutal night of culling thousands of event frames.
← shaped by the manual drudgery that buries the best shots.
🛤 Path
Built a 4-step CV pipeline — label → extract → train → classify (ViT-L + ConvNeXt + LBP) — with a vision-judge grader.
← shaped by proving the pipeline on a real event, end-to-end.
🔀 Pivot
From 'auto-cull' to a human-in-the-loop grader that learns from corrections — Gold / Silver / Bronze, tuned per photographer.
← shaped by taste is personal; the model should learn yours, not impose its own.
💎 Crystal
107 keepers surfaced from a real event; identity tagging + face enrollment — the CV2 image proof-of-concept for the wider platform.
← shaped by one real use-case proving an industry-agnostic architecture.
⭐ Principle
The photographer creates; the machine sorts — and eventually coaches shot-quality live, during the event.
← shaped by augment the human, never replace the eye.