AI CRM

Inspiration

CRMs are generally not useful - I've tried all of them. They require tons of data entry before you can do useful searches, and even then, searching is usually not great.

For example, I want to host an event about curing disease with AI. So who do I know that's interested in healthtech or AI?

Even simpler queries, like "who are the ML engineers", have spotty results.

This app solves both problems.

What it does

We scrape LinkedIn people & company profiles and summarize them with GPT-3.5. We then search over the summaries. These summaries are neutral, of a consistent length, information-dense, and remove hyperlinks. This improves search results.

We minimize data entry by allowing you to join communities like "the Neighborhood" or "The Commons", which are maintained by the organizers of those communities and have large collections of members. You can also create your own, of course. Organizers need only update names and LinkedIn URLs.

I've already found it useful for generating people to invite to my event!

Challenges we ran into

We tried several text embedding and search techniques:

• GPT-3 embeddings don't work well for asymmetric search (where the query is short and the documents, i.e. user profiles, are long)
• OpenAI's new search embeddings worked well for comparing keywords of the same length, but also don't work well for asymmetric search
• Sentence Transfomers did a fine job producing initial search results, but often failed to detect exact string matches

Accomplishments that we're proud of

Trying a variety of text embedding and search strategies, settling on the best one:

• Using inverted index (ElasticSearch-style) to produce exact text matches
• Comparing keywords with OpenAIs new search embeddings, searching for related keywords in the inverted index, and returning all results produced a good initial list
• DistilBart-MNLI did a great job reranking initial results

Designing in 3 hours with Midjourney image-to-image.

What we learned

Generating high-quality profile data with LLMs is half the battle. AI requires rethinking design and programming workflows from first principles! Other takeaways: https://www.figma.com/file/BnwkGPx9cynnZYPLD6FjP2/AI-based-CRM-Presentation?node-id=0%3A1&t=twU2xMYPegm8IZBT-1

What's next for AI CRM

• Implementing the design
• Adding signup
• Getting permission from users and onboarding them
• Adding other data sources, like Twitter, Explorers Club events, or notes from your own private CRM
• Allowing users to edit their generated bios
• Allowing users to suggest to others that they update their bios with their current interests
• Onboarding event organizers

Built With

• faiss
• openai
• streamlit
• tf-idf
• txtai