Inspiration
Exceptional networkers at an event meet ~30 people expecting to make 10 high quality leads in around 2 hours. At an SF Tech week event this means less than 20% of the people on the list. Determined folks go through the list manually and reach out on Linkedin 🥲 - We just don't have enough time to meet who are we looking for at events.
Inspired by latest proposal on generative agents (https://arxiv.org/abs/2304.03442), a concept introduced last year by Stanford’s Park, Joon Sung, et al. can be use to simulate interactions. This interactions uncovers key details that might spark deeper connection ahead of meeting in person.
What it does
Takes in a description of who are you looking to connect at the event. Narrows the list down to relevant profiles. Creates clones of both the relevant profiles collected and thew user. Simulates conversations narrowing down further. Collects insights from conversation. Shares an area to cover when meeting the person at the event.
How we built it
We collect and process attendee data, use matching algorithms, and simulate potential interactions to provide personalized networking recommendations.
Data Collection Scraped Luma app for event attendee lists Gathered LinkedIn profiles for comprehensive user data
Data Processing Created user profiles using an agent-based approach Vectorized data with LlamaIndex Stored in MongoDB vector database
Matching Algorithm Implemented Retrieval-Augmented Generation (RAG) Used cosine similarity to find top three matches
Interaction Simulation Developed multi-agent system (user agent, observation agent, reflection agent, planning agent, conversation agent) Simulated conversations between user agent and match agents Extracted key insights and focus areas for real-life meetings
Additional Features Created pipeline to fetch and store more data as vector database using Pinecone
Challenges we ran into
Implementation of a simplified generative agent memory has been more challenging than expected. This kept the conversations between agents with less details than expected. To mitigate this issue temporarily and showcase the potential we injected more targeted prompts or discussion topics to get useful insights. Iterations over the generative agent arquitecture and ccess to better data (like from companies) should deliver better outcomes.
Accomplishments that we're proud of
We’re proud of building a system that rethinks how matching works by using simulation to dive deeper into potential connections. And honestly, working with such a great team in Palo Alto has been a highlight.
What we learned
LlamaIndex is an absolute game-changer. We learned how crucial it is to move fast and iterate quickly, while making sure the interactions we simulate are meaningful.
What's next for ProNetwork DJ
We want to keep pushing the boundaries, running more simulations to create parallel worlds of "what could happen" during networking. We’re also eager to explore this approach in sales—imagine improving the sales process and lead matching through similar simulations. And of course, we’re excited to keep iterating on this and other projects at Figures. If you have an idea or see an opportunity, let’s grab a coffee and chat! We are looking to expand our web scraping efforts in order to gather extensive user data. This will involve retrieving information from various sources such as Google Scholar, news articles about the person, and any other relevant sources to ensure we have the most comprehensive data possible.
Log in or sign up for Devpost to join the conversation.