We turned 10-Ks, IR releases, Lazard's LCOE+ benchmarks, NREL cost projections, and EIA deployment data into a living, interlinked consulting knowledge base of 67 cross-linked pages using a four-stage process, then hardened it with an integrity layer that fact-checks itself. Here's how it works at a high level.
01 - Extract
We read the filings and reports and identified the knowledge structures inside them: companies across the value chain (Fluence, Tesla Energy, CATL, Form Energy), markets (grid-scale storage, long-duration storage, the battery cell supply chain), and the cost and policy data that moves them. Each source was read once, deeply, and decomposed into atomic, structured fragments - exact figures with their as-of dates, never rounded away.
02 - Structure
We designed a template for each page type - company, market, framework, engagement, source, synthesis - so every fragment follows a consistent schema. A company page always carries its metrics table with fiscal periods; a framework page always carries a worked example applied to real companies in the base. That consistency caught real traps: a fiscal year ending September 30, a headline "profit" that was actually a $220M debt-extinguishment accounting gain, and shipments, deployments, and revenue figures that are not comparable across vendors without a definition column.
03 - Connect
Every fragment was cross-linked into a knowledge graph - 1,060 internal links, all resolving. The Powin bankruptcy links to the vendor-selection framework it motivated, the engagements that applied that framework, and the market shakeout synthesis that puts it in context. Where authoritative sources disagree - Wood Mackenzie's 18.9 GW of 2025 US storage additions versus EIA's ~14 GW - both figures are kept with a methodology note, never silently reconciled.
04 - Compound
This is the consulting pitch made literal: knowledge from past engagements compounds into future ones. The vendor-selection study built after the Powin collapse produced a reusable screen; the market-entry engagement reuses it, plus the cost-outlook engagement's tariff analysis; and a question like "what did our past engagements conclude about integrator counterparty risk" draws across all five engagement pages and files its answer back into the base. Nothing dies in a slide deck.
The integrity layer
A consulting base is only useful if the partner presenting from it can trust every number. Every load-bearing figure - revenue, GWh, $/kWh, market share - carries its as-of date and traces to a source page naming the exact document. Where a figure only exists in trade press (Chinese market shares, private valuations), it is flagged, not asserted; engagement pages state plainly that they are sanitized illustrations built on real public data.
Before sign-off, an adversarial fact-checking pass attacked the base's load-bearing claims against primary sources, and a regression suite of seventeen golden queries runs on every update - if a load-bearing fact silently changes, a red check catches it. The base is auditable and self-checking - it can't quietly rot.
The result: 67 interlinked pages - 15 companies across the storage value chain, 6 market maps, 4 reusable frameworks, 5 illustrative worked examples, 3 cross-cutting syntheses, and 30 source citations - every wikilink resolving (1,060/1,060), 17/17 regression checks green, and the uncertain flagged rather than invented. Compiled and verified by a fleet of AI agents in a day, then hardened - browsable in Obsidian or any markdown reader, and fully owned by the team.