I feel like you guys need to create some additional engineering-level documentation to convey all this. The landing page and few links I navigated to were all too fluffy.
https://people.csail.mit.edu/stavrosp/papers/vldb2017/VLDB17...
The basic jist is: TileDB is substantially faster than alternatives for random writes/reads - this allows for more sophisticated parallel linear algebra algorithms as well as evaluation of predicates.