Tell HN: Selling My SaaS

tl;dr it doesn't.

While I did start with a vision of presbot being a self-learning chatbot built to act as an interactive agent that would represent its owner (primarily b2c) in all sorts of situations. Based on the feedback, I realized that until that interaction is smooth, believable and closer to an actual dynamic conversation, it provides much less value. I was using a combination of https://www.nltk.org/, https://spacy.io/ and https://textblob.readthedocs.io/en/dev/ for NLP then.

I pivoted to a rule-based bot focusing on lead capture via a linear conversation driven by user specified questions, more like an interactive version of a static form, with prescribed Q&A (AKA FAQs on the platform).

Forgot to add, the top level interactions (e.g., lead questions or email capture) require the user to choose an option on the Chat UI (via a button click) to respond, therefore, didn't really have a need to store the context/NLU. I've built some pre-defined responses to open text questions that rely on Levenshtein Distance and threshold (e.g., 90% similarity) for identifying appropriate response. Hope this helps. For stuff that's not defined, the bot says something like "I don't know."