Abuse signals
Automated checks watch play, payment, and registration activity for patterns worth a second look. Everything here is an unreviewed detection — not a finding of wrongdoing.
What the analyst remembers
Two notes travel with every question you ask. The site note is shared with everyone on your site — campaign names, house definitions, conclusions worth keeping. The personal note is yours alone: language, currency, the views you care about, how you like answers written. Write them in plain language, the way you would brief a new colleague. Edit or delete either at any time.
How the analyst works
You ask a question in plain language. A Claude agent plans an approach, looks up the governed data model, writes read-only SQL, runs it against the Redshift warehouse, reads the results, and replies with an explanation and — when it helps — a chart (from trend lines to sankeys and world maps). Every step is streamed live so you can see exactly what it did, and every query is filtered to the sites your account may see.
The pipeline
Question
Plain-language business question.
Understand the model
Searches the curated schema and governed metrics for the right tables.
Write SQL
Drafts one read-only query, preferring canonical KPI definitions.
Run on Redshift
Executes as a read-only user through the Redshift Data API.
Explain + chart
Reads the rows, answers in words, and renders a chart when useful.
Taking an answer with you
Every answer can be downloaded as a PDF (formatted for sharing, with the tables and charts) or as Markdown (the raw text, for pasting into a doc or ticket). Individual result tables also export to CSV.
For slides, just ask — say "turn this into a deck" or "make slides for the board" and the agent writes the deck itself: it picks which findings earn a slide, titles each one with the point it makes, and lays out figures, charts, and tables to suit. The deck appears in the answer with a PowerPoint download.
What keeps it trustworthy
Governed context, not the raw database
The agent only sees a curated, business-facing layer — pre-aggregated marts plus clean
fact and dimension tables, every column documented. It is generated from the dbt project
(model descriptions, allowed values, foreign keys, PII flags), so it never drifts from the
warehouse. The raw platform_core and staging tables are never exposed.
Governed metrics
Canonical KPI definitions (GGR, NGR, active players, …) live in the dbt model metadata and are handed to the agent verbatim, so it uses the same numbers the BI team does instead of re-deriving them from raw events.
Read-only by construction
Every query passes a guard before it runs: a single SELECT/WITH
statement only — no writes, no SELECT … INTO, no LIMIT ALL — and a
row limit is added automatically. It connects as a dedicated read-only warehouse user, not an
admin. The same core is also exposed over MCP for other clients.
Privacy
Columns that hold personal data are flagged in the model and the agent is told to prefer aggregates over listing individuals.
Who sees what — roles, enforced in SQL
Every account is a principal with a role. Admin sees all data and manages
accounts from the Admin tab. Platform sees all data. Operator sees every
site belonging to its operators. Site sees only its named sites (a BetMode
account sees BetMode). For operator and site roles, every query the agent writes is
rewritten server-side so each table is filtered to the allowed site_ids —
tables that cannot be filtered are blocked, and the live catalog is hidden. The filter
is enforced in SQL, not by the model's goodwill.
What the analyst remembers
Two notes, carried between conversations
The site note is shared team knowledge — campaign names, house definitions, conclusions from earlier analyses — visible to and editable by every account on your site. The personal note holds your own preferences (language, currency, favourite views) and is yours alone. Read, edit, or delete both on the .
Chart types the agent can draw
Tools the agent can call
Governed metrics
Data dictionary
Business glossary & SQL dialect
Loading glossary…
Accounts & access
Create an account per audience, pick its role and sites, and hand over the generated token. Site and operator accounts are filtered in SQL on every query; disabling an account or rotating its token takes effect within seconds.
Create account
Accounts
Spend limits live in the usage panel — the site's monthly cap and a per-account cap for anyone here. Pick the site at the top of the panel to change one.
Usage & budgets
Month-to-date AI spend per site: $100/month included, then automatic metered overage billed at cost + 15% on the operator's monthly settlement, up to the site's own spend limit (self-service, default $500). Accounts inside it are uncapped until a site admin caps them. Pick a site name to set either.
From server configuration
Principals defined in the AnalystPrincipals secret. Read-only here — change
the secret and redeploy to edit them.