Boundless Leader: an auditable matching engine built on Claude

When a leader hits an inflection point, finding the right guide should not be left to chance. Yet the default is a directory: rows of profiles, no real sense of fit, and a decision made on guesswork at exactly the moment it matters most. Boundless Leader replaces that with a conversation designed to understand the person first.

The architecture is a staged pipeline behind an agentic loop. Claude runs the assessment as a genuine dialogue — deciding when to clarify, when to re-ask, and when to advance — driving progression through inline structured markers and calling its own tools to present options, trigger matching, and close the session. The result feels like a conversation, not a form, because that is exactly what it is.

Once the conversation completes, four specialized Claude stages run in sequence, each with its own model tier, tool schema, and prompt. Profile extraction reads the user's full answer pattern into a structured leader profile. Scoring applies per-criterion judgment across the practitioner pool. Matching writes the user-facing rationale for each recommendation. And a narrative stage generates the consequence-and-opportunity prose that makes the match land.

Every one of those stages uses forced tool choice with strict JSON schemas, so Claude returns validated structured output rather than prose that has to be parsed. The weights, ranking, and tie-breaks stay in deterministic code. That split is deliberate: it keeps selection auditable and reproducible while the model does the part only a model can do — reading nuance and writing judgment.

To keep the experience fast and cost-efficient at scale, Boundless relies on prompt caching across both the system prompt and the growing conversation history, response streaming with mid-turn recovery, strict tool use, and per-tier model routing. It all sits behind a provider-agnostic abstraction with full observability on tokens and cost per stage.

The result is a real, mature use case of Anthropic's technology as the core product layer — not a support chatbot bolted onto a directory. Claude runs the conversation, extracts the profile, scores the pool, and writes the rationale, while deterministic code keeps the outcome auditable end to end.