Build with DataHub: The Agent Hackathon · Production ML Agents

Recall

A lineage-native pre-mortem agent for production ML — it finds the upstream metadata failures that silently corrupt models, before the predictions go wrong.

Recall walks every production model's full upstream closure in DataHub, detects the structural failures that don't throw exceptions, scores each model 0–100, and writes the findings back into the graph as assertions, incidents, risk properties and a human-readable pre-mortem — so the next engineer or agent inherits the diagnosis instead of repeating the investigation.

Start here — real output from a real run

Before · healthy estate

Baseline scan

0Risk, both models
8Baselines set

Every model HEALTHY. First sight of a dataset reports BASELINE_ESTABLISHED, not drift — there is nothing to compare against yet.

Open the report
After · five failures injected

Pre-mortem scan

3Critical
2Warn
100Top risk score

Every injected failure caught, each traced to the exact dataset, feature and training run responsible — with agent-written remediation.

Open the report

Both are self-contained HTML — no server, no build, works offline. Everything on this page is generated by scripts/generate_examples.py against a live DataHub, so it can't drift from what the code actually emits.

What the second scan found

fraud_detector_v3 100 CRITICAL 2 live deployments
churn_predictor 14 WATCH 1 live deployment

The detectors that fired

Orphaned feature orphaned_feature

A feature still declares a soft-deleted dataset as its source. The table is out of the catalog; anything still writing it is outside governance.

Training/serving skew training_serving_skew

A feature served to the model derives from a dataset the training run never consumed. The model is scoring against data it was not fitted on.

Schema drift training_source_schema_drift

A dataset backing a live serving feature changed shape since the recorded fingerprint — columns added, dropped, or retyped.

Ownership vacuum ownership_vacuum

A dataset feeding a production model has no owner. When it breaks there is nobody to page.

Blast radius blast_radius

One shared upstream dataset sits beneath multiple production models — a single change lands across several deployments at once.

Pre-mortems written back into DataHub

Each is stored as a DataHub document linked to the model and every implicated dataset. The remediation was written by a Claude agent that investigated the graph through the DataHub MCP server first — reading deprecation notes, diffing schemas and looking up owners, rather than paraphrasing the finding.

Machine-readable output

Run it yourself

# DataHub + the healthy estate
make up && make seed

make scan     # baseline — everything HEALTHY
make break    # inject five realistic failures
make scan     # every one of them caught
make reset    # back to healthy

Add --narrate for the agent-written remediation and --html out/report.html for the rendered report. Full walkthrough in TRY_IT.md.