- Caught
- In CI, before the merge. Verdict posted as a PR comment.
- Check
- After transformation:
orders_enriched - Finding
- Column
customer_idwas renamed toclient_id. 2 downstream joins now drop 18% of rows. - Why it matters
- Revenue by customer on the dashboard will be understated by ~18%.
- Suggested fix
- Alias
client_id AS customer_idinorders_enriched, or update the joins inrevenue_by_customerandchurn_daily. - Changed by
- AI agent (session #4821)
The agent that checks your agents' data pipelines.
Know your data is right before anyone sees it.
AtEvals checks every change: at ingestion, after each transformation, and on your dashboard KPIs. Then it decides: publish, warn, or hold.
Scenario 1 of 3: All clear
Hover or tap a check to see what it looks at.
WHY ATEVALS
Your tests passed. The dashboard is still wrong.
Your tests
- Written ahead of time, for cases you predicted
- Check your transformation logic, not the source system
- Pass even when a source sends bad data (e.g. ERP or CRM)
- Pass/fail with no context
AtEvals
- Checks the data your pipeline actually produced
- Catches issues nobody wrote a test for
- Flags wrong KPIs, even when the source system is at fault
- Traces the cause and says what to do next
Keep your tests. AtEvals checks what they can't see.
UNDER THE HOOD
Your agent calls AtEvals. AtEvals checks the data.
Your coding agent
Any MCP-compatible agent
AtEvals
- check_model
- check_kpis
- explain
①Warehouse
sources & tables
②dbt project
the PR diff & models
③BI tool
KPI definitions
agent → atevals.check_model("orders_enriched")← HOLD customer_id renamed · 2 joins affectedagent → fixes join in revenue_by_customeragent → atevals.check_model("orders_enriched")← PASS 14 checks passed · ready for reviewAtEvals ships as an MCP server, so your agent checks its own work before handing off, or as a step in your CI.
How it works
Three checkpoints, raw data to dashboard
01 · INGESTION
When new data lands
Schema changes and stale sources
- Columns added, dropped or retyped
- Late or missing loads
- Unusual row counts
- Null spikes on key columns
02 · TRANSFORMATION
On every PR, after your tests
Broken logic your tests missed
- Diff against production
- Joins that drop or duplicate rows
- New duplicates or nulls
- Totals that don't match upstream
03 · KPIs
Before dashboards refresh
Wrong numbers, healthy pipeline
- KPIs read from your BI tool
- Out-of-range values and spikes
- Root cause: pipeline or source (ERP)
When it runs
Two moments. Same checks.
Before the merge
Catch it while it's still a diff.
Runs on the pull request. Shows which models the change touches and what would break downstream.
Verdict
A comment on the pull request: pass, warn, or block the merge.
Every pipeline run
Catch it when the data changes underneath you.
Runs on every scheduled pipeline run. The only place a source-system problem shows up, like an ERP that exported the same orders twice.
Verdict
Hold publication and name the cause.
The merge catches what your change broke. The pipeline run catches what the world changed.
Example finding
Not just a report. A verdict.
Source shopify_orders last updated 26h ago (expected: hourly)
14 checks passed on stg_customers
Get early access.
What AtEvals can see
- Read-only access to the schemas and models you choose
- Your raw data never reaches the AI model: only metadata and check statistics do
- AI models and processing hosted in Canada
- Stores check results and aggregate metrics, not your raw rows