Inspiration

Most broken dashboards start as a small upstream change: a column switches from dollars to cents, an ID gets merged. Data contracts describe what "correct" means, but they're written in English and rarely turned into tests.

What it does

lgnemo check stg_orders.amount --change "switch to integer cents"

  1. Walks the lineage graph to find every model and dashboard downstream, with the path that explains each one.
  2. Collects the data contracts on those paths.
  3. Nemotron writes a SQL check per contract, runs it in a read-only warehouse sandbox, reads any error, and repairs the SQL.
  4. Outputs a ship/hold report, who to notify (on-call first), and the checks as .sql files ready for CI.

How we built it

  • NVIDIA Nemotron 3 (Super nvidia/nemotron-3-super-120b-a12b on Nebius Token Factory for the live demo; Ultra nvidia/nemotron-3-ultra-550b-a55b via NVIDIA's API as the alternate host) via the OpenAI-compatible chat completions API with tool calling. Model size is a flag (Nano for quick checks, Ultra for big graphs).
  • Four grounded tools: impact, owners, warehouse_schema, run_check.
  • Python stdlib only: in-memory lineage engine, SQLite warehouse with an authorizer that blocks writes.
  • Synthetic warehouse with planted defects that replay real-world incident patterns.
  • Fifteen pytest tests + GitHub Actions CI.

Challenges we ran into

Keeping the agent honest: every claim must come from a tool result, and SQL runs read-only so a generated query can't damage data.

Accomplishments that we're proud of

The write → run → repair loop works end to end. Live on Nebius Token Factory, Nemotron Super traced the 36-asset blast radius of stg_orders.amount, wrote and ran 5 SQL contract checks, caught the planted cents-vs-dollars defect (2 rows over the contract's 10,000 USD ceiling via test:accepted_range_stg_orders_amount) and recommended HOLD in about 34 seconds (8 tool calls). On the alternate NVIDIA host, Nemotron Ultra runs also catch both planted defects: cents-vs-dollars via stg_orders.amount and the duplicated customer_id via stg_customers.customer_id.

What we learned

  • Nemotron 3 Ultra (nvidia/nemotron-3-ultra-550b-a55b) drove the full tool loop live with zero parsing changes: clean OpenAI-style tool_calls, valid JSON arguments, sensible tool order (impact → schema → checks → owners).
  • Agents need guardrails on coverage, not just syntax. On the first live run the model checked look-alike dbt tests and skipped the governing contract: ids. A one-shot "these contracts are still unchecked" reminder in the loop fixed it (now unit-tested).
  • Thresholds matter: with a loose range the model passed bad data; telling it to flag implausible magnitudes for the unit let it catch the planted cents-vs-dollars defect (2 rows) and recommend HOLD.
  • Its reasoning was good: it noticed the raw column is already total_price_cents and warned that a partial rollout inflates revenue 100×.
  • Grounding coverage in the graph helps: once impact listed the tables each contract governs, the reminder could name exact gaps (e.g. mart_customers), and the model then checked them. Allowing two reminders instead of one made contract:orders_amount catch the defect on stg_orders. Stating the 10,000 USD ceiling in the contract spec also beat letting the model pick a threshold.
  • Ready-made rules close the last gap: impact now hands each contract its machine-checkable clauses as SQL (not-null, accepted range parsed from the spec, plus the dbt tests that enforce them), and the coverage reminder names any rule whose bounds no check has used yet. With that, Nemotron Super on Token Factory ran the range test with the contract's own 10,000 USD bound and caught the 2 bad rows.
  • The duplicated customer_id was caught live (2 rows) on the stg_customers.customer_id run in about 2 minutes.
  • Expect transient endpoint hiccups (one read timeout, one HTTP 500): the client now retries with backoff.
  • Latency: a full run takes about 34 seconds on Nemotron Super via Token Factory (8 tool calls, ~46k tokens), and 2–5 minutes on Ultra; both fit a pre-merge check.

What's next

dbt manifest import, emitting dbt tests directly, GitHub PR comments on schema changes.

What was significantly updated during the Submission Period

New repository created Oct 28, 2026. It reuses the idea and the synthetic dataset from my earlier LineageGate project; the coding-agent loop, SQL sandbox, warehouse generator, Nebius/Nemotron provider, CLI and tests are all new.

Built With

Share this project:

Updates

Submission history