AI AGENTS
Re-check generated Replicate images and quarantine unsafe output
After Replicate finishes an image, this workflow runs the result through a HuggingFace image-safety classifier, deletes or quarantines unsafe images.
How it runs
The automated pipeline, trigger to output.
- TriggerReplicate prediction completes (webhook)Replicate
- ActionClassify output image for safetyHugging Face
- LogicBranch on unsafe score vs threshold
- ActionMove unsafe image to S3 quarantine bucketAWS S3
- ActionLog image reference and disposition to PostgresPostgres
- OutputReturn released URL or quarantine noticeHTTP webhook
What it does
Prompt screening catches unsafe text, but generators can still produce unsafe images from innocent-looking prompts. This workflow closes that gap by classifying the actual output image after generation. Safe images are released to the user; unsafe ones are quarantined in object storage and flagged, never delivered.
When to use it
Use it as a second layer behind a prompt gate, or on its own when your risk is in the imagery rather than the wording. Essential for public-facing generators where a single unsafe image is a serious incident.
How it works
- 1A Replicate prediction-complete webhook delivers the finished image URL.
- 2The image is run through a HuggingFace image-classification safety model.
- 3A logic branch checks the unsafe score against the threshold.
- 4Unsafe images are moved to a quarantine bucket in AWS S3 rather than served.
- 5The image reference, scores, and disposition are logged to Postgres.
- 6The caller receives either the released image URL or a quarantine notice.
Set it up
What you configure once, before turning it on.
- 1Connect ReplicateImage, video, and model inference.
- 2Connect Hugging FaceModels, datasets, spaces — the open-source hub.
- 3Connect AWS S3Buckets, objects, signed URLs.
- 4Connect PostgresAny Postgres URL — query, write, migrate.
- 5Connect HTTP webhookTrigger any URL on agent actions.
- 6Set each agent's modelWe leave models unset so you pick the tier — fast + cheap, or top-quality.
- 7Tune it to your dataEdit the prompts, filters, and field mappings so it matches how your team works.
- 8Test, then turn it onRun once against a sample, confirm the output, then enable the trigger.
More AI Agents workflows
Observability Cost Allocation Report
Monthly, an agent pulls Datadog and Honeycomb usage, allocates spend to teams and services by tags, writes the breakdown to Snowflake, and posts a chargeback summary to Slack.
Vendor Shortlist Matrix from a Buying Brief
An agent reads a buying brief, researches candidate vendors across the live web, and builds a scored comparison matrix in Coda ranking each vendor against your stated criteria.
Split oversized Linear epics into estimated child issues
An agent scans newly created Linear epics, breaks each one above a size threshold into discrete child issues with point estimates and acceptance criteria.
Datadog Bill Spike Attribution Agent
When a daily Datadog cost check detects a spend jump, an agent attributes the increase to the specific services and metric types driving it and posts a ranked breakdown to Slack.
Buying Brief Email to Shortlist Doc in Drive
When a buying brief arrives by email, an agent researches the market and produces a polished narrative shortlist document in Google Drive, then replies to the sender with the link.
Zoom Demo Low-Score Objection Escalation to Manager
Scores how well a rep handled objections in each Zoom demo, and only when the handling score falls below a threshold does it create a coaching task in ClickUp and alert the rep's…
Run it inside a business
This workflow drops into a full company template. Import the org, and this is one of the playbooks its agents run.
Run this workflow in your colony.
14-day trial. No DevOps. No Sales call. Provisioned in under a minute.