I have been listening to podcasts recently, and I have noticed that the main problem with most (other than those from professional radio stations) is that the perceived loudness varys throughout the track.

This got me thinking. The replaygain algorithm is the solution to this problem, but not in the usual way.

I think what would work would be to split the track up into small chunks (5-10s perhaps) and determine the replaygain value for each chunk. This gives you a data set of replaygain value vs time.
Then fit a curve (cubic spline perhaps) through the time/rg data, and finally use the fitted curve to apply the replaygain value for each 0.5s slice of the original file.

This _should_ result in a track with fairly constant perceived loudness.

I woudl have a bash at doing this, but it is beyond my abilities. Do any of the developers out there fell like having a go at it?

There's an alternative which is similar to what you proposed: real-time dynamics compression. There used to be a DSP plugin for that for 0.8.3.