Running a complete example with TIDL

Hi,
I am responsible for investigating TIDL for our company, If you could help me evaluate TIDL correctly, I would appreciate it.

I have been investigating model generation, TIDL import, and TIDL usecases with VSDK 3.04.
I have successfuly ran all the usecases -with the available binary examples- which are tidl file /IO, tidl semSeg, tidl OD. They work on the board.
I have successfuly generated binaries for three example models: inception v1, mobilenet, jdetnet, but they don't work on the board.

Status
My focus is to have one complete example from Tensorflow/Caffee code or model -> to Binary Output -> to having it running -relatively- correctly on a usecase. (preferrably object detection with SSD)
There exists this link (https://e2e.ti.com/support/processors/f/791/t/735984
but sadly I wasn't able to verify the output running on actual hardware. Sometimes there is no image or sometimes the SSD boxes are meaningless.

1) I am wondering if some example exists such that I can generate and train the model and see the output from the usecase directly?
TensorFlow/Caffee Code --> .ckpt or .proto --> TIDL Import tool -> net.bin and param.bin files -> USECASE
This is very important since you can make sure an example works on the board and can focus on your model training instead of all the intermediary steps.

2)Is such a workflow/usecase verified with the open source resources,-other than TI's hidden dataset-? I can not use any of the TIDL training examples with the available usecases.
The answer to this question would reshape our understanding of TIDL and influence our idea of using TDA2xx for our future products, so it is very important to us.

3)Are usecases generic, -with simple modifications to labels and usecases-? Imagine is one example for semantic segmentation, there is one for classification, and there is one for object detection.
Can I use any model with the same purpose to expect a viable output?

4) Please help us understand what the usecases and algorithms actually do. What are the algorithm outputs, and are they standardized? A documentation would be nice, saying that our object detection algorithm is responsible for doing <THESE> It gets <THESE> inputs, it outputs <THESE> You can configure it by using <THESE> It is compatible with <THESE> usecases. It is compatible with <THESE> models and datasets.
Investigating the user guide, I feel like TIDL relies much on the community information that is scattered around rather than TI official guidelines, that brought some challenges for me to understand.

For example, it would be great if there was a verified usecases table, that one could pick up and test - from the code or pretrained model to using it in a usecase-

We'd appreciate if you could help us with our doubts. Thank you.