Chive.
c. 2013
Talent intelligence that learns from judgment.
Chive was a talent intelligence system developed through independent consulting for venture capital firms. It connected changing signals about people and organizations with the judgment of the researchers using it, helping them decide where to direct their attention and when to begin a conversation.
Relevance and timing.
Someone can be an exceptional prospect long before the right moment to approach them. Chive treated relevance and readiness as distinct questions. Career moves, new connections, and profile changes supplied evidence that could be evaluated alongside a researcher’s knowledge of the person and the opportunity.
The data model connected people, organizations, and groups through histories of activity. Features captured changes across multiple time horizons, allowing recent activity and longer-term patterns to inform the same decision.
A model for each point of view.
Each researcher and research question had a separate model. Feedback recorded which suggestions were useful and supplied examples for training. The learning system could reflect the priorities of a particular search and the judgment of the person conducting it.
An Erlang service managed the models, extracted features from event histories, and ranked candidates. Feedback from the interface triggered retraining, connecting the act of evaluating a suggestion to the system that produced the next set of recommendations.
Preserving the context of a decision.
A judgment belongs to a moment in time. Someone’s later career move should not become evidence for a prediction made before it happened. Training examples carried timestamps, and their features were reconstructed from the history available at that point. This kept the learning process tied to what could have been known when the feedback was given.
Keeping research in motion.
The product brought together data ingestion, event storage, contextual browsing, feedback, and personalized rankings. Researchers could inspect a person’s history, follow relationships, and refine suggestions within the same workflow. Scheduled email digests brought their highest-ranked prospects back into view.
The work carried an open-ended consulting problem through data modeling, machine learning, backend engineering, and interface design. Human judgment became a continuing input to the system, giving the research process a way to retain and apply what it learned.