AlchemyLake Is Live, Truth Made Visible
Pull-quote: “The model never authors a number. It writes around figures that were computed by code, and after it writes, a verifier checks every numeric claim against the facts table and scores the result into the seal.”
The announcement
AlchemyLake is live at app.alchemylake.com. It is a governed creative platform: it takes the data your lakehouse already trusts and produces the deliverables your organization actually ships, with the numbers computed by code, verified after generation, and sealed to their source.
Enterprises solved data governance and then handed the last mile to ungoverned tools. The chart in the board deck was retyped by hand. The infographic number came from a prompt. The narrated video came from nowhere auditable at all. AlchemyLake exists to close that gap, a category we call governed creative.
What is live today
Eight studio lanes. Analyst chat, deep research dossiers, enterprise PDF reports with statistical appendices, presentations with real charts and speaker scripts, designed infographic posters, animated video briefings, two-host podcast briefings, and data-driven music scores. Every lane binds to a governed source before it renders anything.
Governed sources, four ways. Databricks Unity Catalog tables under an explicit allowlist, Genie spaces whose conversations become bindable sources, uploaded CSV and Excel workbooks, and unstructured documents that bind as a cited corpus. A built-in sample lakehouse lets anyone try the full pipeline in minutes.
Numbers discipline. Before any model writes a word, a deterministic facts engine computes the figures: totals, deltas, trends with fit quality, outliers, correlations, concentration measures. The model composes narrative around fact tokens. It cannot invent a number, because it never holds the pen on one.
Verify and seal. After generation, a numbers-fidelity verifier extracts every numeric claim from the deliverable and checks it against the computed facts. The verification score, source identity, row count, and content hash are embedded in the file itself, PDF, image, audio, or video, so provenance survives download and forwarding.
Agent-native from day one. The same engine is exposed through thirteen MCP tools, a public REST API, and a CLI. An agent can ask Genie a governed question and render the board deck in a single turn.
The pipeline in one line
Bind, analyze, design, transmute, verify, seal. Six stages, and the two that matter most, verify and seal, are the two the industry usually skips.
AlchemyLake is live with a free starting grant of credits and no card required. Try it at app.alchemylake.com, or read how it complements a Databricks estate on our Databricks practice page.