I've done a similar architecture as well a couple of times. Both setups use multiple haproxy instances (auto-scaling group), multiple backend app servers, and each haproxy server has a local nginx instance.
Nginx is used for serving some static error pages that live outside of the backend applications haproxy talks to (eg: "Customer domain not recognized"). There's some hacks to kind of make this work in haproxy but it's much simpler with an actual HTTP content server.
Nginx also works well as a caching proxy, and what's really cool with haproxy is it's easy to have most requests go directly to the application server (avoiding an unnecessary proxy hop), but just pattern match certain types -- eg regex match `.jpg$` -- to go through the nginx cache.
I've also used a similar technique to have per-domain custom images overriding specific URL paths (eg `/logo.png`), which was a really quick way to allow some customization of a multi-tenant application without having to make major code modifications to support it (only used for a tiny handful of customers).