# Answers for 30+ Teammates > How a versioned GitHub atlas replaced word-of-mouth warehouse knowledge — so 30+ teammates and their agents could stop chasing the person who knew. - Author: Paul Yu - Role: Senior Analytics Engineer - Employer: Viral Nation - Year: 2026 - Theme: Documentation - Human article: [https://paul-yu.com/case-studies/data-knowledge-and-lineage](https://paul-yu.com/case-studies/data-knowledge-and-lineage) - Structured data: [https://paul-yu.com/agent/case-studies/data-knowledge-and-lineage.json](https://paul-yu.com/agent/case-studies/data-knowledge-and-lineage.json) ## Evidence | Value | Measure | Evidence level | | --- | --- | --- | | 466 | warehouse objects indexed | Repository backed | | 13,853 | physical fields catalogued | Repository backed | | 37.2K | original documentation lines | Repository backed | | 89 | versioned commits | Repository backed | ## Situation Hundreds of warehouse tables existed, but the rules that made them safe lived in meetings, chat, and a few humans’ heads. New teammates had to chase whoever happened to know. ## Task Turn scattered data and domain knowledge into one maintained layer that could onboard both humans and agents without retelling the whole system. ## Action I built a versioned GitHub atlas that combines generated warehouse inventories with curated mental models, object contracts, lineage, source SQL, metric rules, caveats, and guided learning paths. A deterministic builder packages the same source files for agents. ## Result More than 30 data team members now share one reference spanning 466 objects and 13,853 fields. New teammates and agents start from the same map instead of rebuilding it from conversations. ## System flow Warehouse objects → Lineage + contracts → Searchable knowledge → Safer use ## Technical notes ### Data lineage graph The lineage map follows data from its source, through each transformation, to reports and AI tools. If a number looks wrong, the map shows which layer owns it. ### Data contract A contract records the row grain, keys, fields, caveats, and expected behavior. Producers and users can review the same promise when something changes. ### One owner per rule The atlas is not one giant table. Each rule has one home, and other pages link back to it rather than keeping their own copy. ## Ownership I designed the atlas, curated its data and domain knowledge, wrote its main contracts and lineage guides, built its metadata snapshots and agent bundle, and maintained all 89 recorded commits. ## Limitations The repository proves coverage and maintenance. The 30-plus users are a direct operating observation; there is no page-view log or controlled onboarding-time study. ## Takeaway Data expertise scaled because documentation became infrastructure. Generated metadata covers the warehouse; curated guidance carries the meaning; Git keeps humans and agents on one reviewable map. ## Tools GitHub, Markdown, Python, Snowflake, BigQuery, Data lineage