At scale, the weird record that signals fraud or a broken feed looks identical to the millions around it. This flow finds the outliers by similarity, not rules.
sonnet1 weekPineconePythonscikit-learnKafka
Claude
53ROI
71Scale
$2.2k84Saved
ROI for
README.md
Why this subagent
At scale, the weird record that signals fraud or a broken feed looks identical to the millions around it. This flow finds the outliers by similarity, not rules.
Each record is embedded and indexed; a KNN lookup measures how far it sits from its neighbors, and anything past the threshold gets flagged. Flagged points arrive with their nearest neighbors attached, giving a downstream agent the context to judge whether it's a real anomaly.
How it runs
Used at step 01 to kick off the pipeline.
Write
Used at step 01 to kick off the pipeline.
WebFetch
Used at step 01 to kick off the pipeline.
WebSearch
Used at step 01 to kick off the pipeline.
Run a KNN lookup to find each new point's nearest neighbors in the existing space.
pending
Flag points whose distance to their neighbors exceeds the learned threshold as anomalies.
pending
Emit flagged records with their neighbor context so an agent can ground its review.
pending
Sample output
json
// Sample output
// (generated when the pipeline finishes)
Embed each record, run KNN, and flag points whose neighbor distance exceeds the threshold, with neighbor context attached.
Unlock the rest
The full agent definition, install snippet, and starter task are gated for community members.
Members get the full `.md` agent file, the npm / pnpm install one-liners, a starter prompt that we've tuned against real runs, and the open-source repo when this automation ships there. One email, magic link, done.