I'm actually using the algorithm that Evan Miller writes about (i.e. the lower bound of a binomial proportion confidence interval), in the discussion system that powers the demo in the linked article. I'm using the Agresti-Coull interval, http://en.wikipedia.org/wiki/Binomial_proportion_confidence_..., however, because it seemed easier to implement. (I didn't know about Evan Miller's article (or any other).)