Team knowledge, finally connected

The setting

A SaaS product serving hundreds of thousands of users for everyday team updates. Users report their work and can also see the work of others on their team.

The challenge

Extract organizational knowledge from daily user updates. Identify potential blockers and corresponding actions, team well-being, important team members, organizational silos, task completion, and more. Provide a fully conversational AI assistant that makes this knowledge available to anyone.

Why a custom solution

Much of what was needed had never been built before. Two of the methods were new: a model of team well-being as a time series, based on metrics from organizational psychology, which resulted in US patent 20250061406-A1; and a topic-modeling approach that combines BERT sentence embeddings with network analysis.

The work

A team knowledge graph was constructed in Neo4j. Its entities included users, reports, tasks, blockers, actions, and more, while its edges represented the relationships between them. The graph was populated through ETL pipelines orchestrated with Airflow, with the data transformed by a series of NLP models and classifiers. Figure 2 shows some of the information extracted from user reports, while Figure 1 shows an example organizational network for a team. Additional algorithms handled specialized tasks, including temporal task recognition using graph algorithms and prediction of user reporting times.

The results

• A fully conversational AI agent that assists team members.
• 85% accuracy in predicting user reporting times.
• A team well-being model with 95% accuracy, presented as a time series.
• US patent 20250061406-A1 for the team well-being model.
• Advanced internal analytics for churn prediction, user engagement, product adoption, and growth prediction.
• Dashboard with customer analytics, insights and forecasts.
• An audience generator model for targeted marketing campaigns.
• Topic modeling using a novel approach that combines BERT sentence embeddings with network analysis methods.