This raises an obvious question. Why Gleam team weren't doing that in the first place? Obviously they didn't have direct Gleam AST -> Erlang source pipeline, they must have had to build some their own custom Erlang AST representation that they then translated into the source code.

We did not (and still don't) have an internal Erlang AST representation in the compiler. Previously we did have a Gleam AST -> printing algebra -> Erlang source pipeline, now we construct a buffer of bytes with the data encoded as Erlang Term Format.

Source was the previous target as at the time Erlang Abstract Terms was not established as the go-to format (Core Erlang was more popular but it did not have a stable API outside of the BEAM, so Gleam's in-Rust compiler could not construct it), and due to the newness of the language having an "escape hatch" where one could abandon Gleam and eject to Erlang was highly valuable. It also meant we could use the Erlang build tool until the Gleam one was ready for use.