⌘K
API referenceMCP Server

Architecture

How data moves from source to a ranked, evidence-backed answer.

Data in
  • ATS, HRIS and LMS
    Connected systems, kept in sync
  • File upload
    Resumes, job descriptions, org data
  • SDK and API
    Automated feeds
Structuring
text becomes facts
The Graph
Occupations, skills and how they relate
plus each company’s people, roles and Success Profiles
Talent Intelligence
rank, explain, audit
Answers out
  • LLM Connectors
    Claude, ChatGPT, Gemini
  • MCP Server
    AI agents and coding tools
  • WIC
    Work Intelligence Console
  • API and SDKs
    Python, Node.js, Java
Everything goes into the Graph, and every answer comes out of it. Data enters from ATS, HRIS and LMS integrations, file uploads, or the SDK and API; structuring turns it into facts on the Graph; and Talent Intelligence ranks, explains and audits the answer that every surface — LLM Connectors, MCP Server, Work Intelligence Console, and API and SDKs — returns.

How to read this

Data enters Workerbee from a live integration, a direct upload, or the SDK. Before anything is matched, it passes through structuring, which turns free text into normalized fields grounded against the Graph — the shared vocabulary of occupations and skills that makes “Cloud DevOps Engineer” and “Senior Cloud Infrastructure Engineer” comparable in the first place. It’s built from hundreds of thousands of job descriptions and resumes, in addition to public taxonomies from O*NET, ESCO and Lightcast.

Talent Intelligence sits on top of the Graph: it’s what turns a structured profile and a structured role into a ranked, evidence-backed answer, with an audit record attached.

Every surface — the REST/MCP API, the MCP Server for AI agents, the Work Intelligence Console (WIC), and the LLM connectors inside Claude/ChatGPT/Gemini — calls the same Talent Intelligence layer underneath, so an answer is consistent no matter where the question was asked from.

Where the model sits

Workerbee uses an LLM at two points only: reading source data during structuring, and phrasing the answer to a question. Everything in between — the Graph, Success Profiles, matching, and ranking — is structured data and math, not a prompt.

That’s why answers hold steady when you switch providers or a model is upgraded. The standard and the rankings live in the Graph, so a new model changes how an answer is worded, not what it says — and nobody has to re-run or clean up past work after an upgrade.

Where the cost is

Workerbee’s cost sits at ingestion: data is structured onto the Graph once. After that, a question is a lookup against structured data, not a fresh model read of raw documents. Adding users or asking more questions doesn’t multiply cost the way sending full context to a model on every question does.

Your data boundary

Each customer’s data is scoped to its own tenant. Cross-customer intelligence, such as market benchmarks, is limited to anonymized aggregates — individual records never cross tenant boundaries.

What stays out of the Graph

The Graph’s structured schema has no fields for protected-class attributes, so that data isn’t collected by design. Changes to a Success Profile are proposed first and only take effect once someone approves them.

For authentication, data residency, tenant isolation, and audit detail, see the Technical & security overview.

Last updated