[abannin]

I suspect the issue is one of two scenarios: 1) CEO has low bandwidth, hands off a nice-to-have-but-not-critical feature. They don't have to invest much time; if it works, awesome, if not, no problem. If you don't complete the feature, it may or may not get done. 2) Company is overwhelmed, feature is perceived to be easy(-ish). It will get done, you're just a first pass.

If (1), you're probably best served not asking for much help. If (2), you need to educate & manage expectations about the complexity and necessary steps for the feature. Either case, you would probably be best served by prioritizing the input/output functionality (input a picture, output a prediction) over optimizing precision/recall on the algo. This allows you to 1) show progress to others and 2) you can start having conversations about false positives/negatives by showing different images and the prediction output.