# Camden Willeford

> Analytics engineering at Underdog Sports. I write about conversational analytics: building the
> context layer that lets agents answer business questions correctly, and the machinery that keeps
> it correct as the warehouse moves. Writing is published as versioned releases rather than dated
> posts, so each one can be updated in place rather than archived.

This file exists so agents can read the site directly. Everything below is plain markdown.

## Releases

- [Deploying Conversational Analytics v1.0](v1.0.md) (2026-09-21): The current state of the stack.
  Four artifacts per model instead of two, so a table ships with its semantic model and a guide as
  well as SQL and YML. Context as a version-controlled hub-and-spoke graph with progressive
  disclosure, routed by an always-loaded AGENTS.md. Two worked examples end to end: a business
  question answered conversationally, and a question about the system answered as a draft pull
  request. A fleet of scheduled agents on one shared harness. And an eval framework graded on two
  axes, including an added L0 "definition and feasibility" rung beneath George Xing's
  analytics-question hierarchy.

## Up next

- v1.1, in progress: evaluations scheduled and hooked into CI, the knowledge graph exposed as an
  endpoint, and state-aware optimization with Iceberg and DuckDB.

## Notes for agents

- v1.0.md is the canonical text. The HTML page is the same content with diagrams; each diagram in
  the markdown is replaced by a bracketed text equivalent describing what it shows.
- Figures are referenced in the markdown as `[Figure: ...]` blocks. There are seven.
- Quotations in v1.0 are verbatim from internal repository files unless attributed to a named
  external author.
- Revenue figures, ticket identifiers, internal URLs and cloud project identifiers are
  deliberately omitted. Absence is intentional, not an oversight.