If you don't understand the value of TileDB, you probably haven't dealt with the data that it is meant to model.

Using something like MongoDB for genomic or sparse geospatial data is incredibly cumbersome.

The access patterns for analytical applications making use of the above data types are spatially collocated ... the subsequent implication of this is that data needs to be stored in ways that are geometrically optimized on the file-system.

Consider a dataset which assembles information regarding terrain ... you collect a measurement using a laser. Some areas you use a laser scanner that is very dense (high number of samples per square km) ... in another you sparsely sample it. This results in a huge multi-dimensional array... this is well understood by people collecting the data. TileDB is built for this type of use case... "complex" data is ambiguous.