This looks so really cool, we tried to build something similar in-house for our SaaS product but it took too long and wasn't reliable enough. Would love to retry with superflows sometime!
Totally get the reliability issues you've seen - getting LLMs to output the right response 80% of the time on these kinds of tasks is not that hard, the last 20% is the real challenge.
On trying us out - ping me an email at henry@superflows.ai