WardenWarden

Community platforms need to stop harmful posts without making every user wait for a model.Community platforms need to stop harmful posts without making every user wait for a model.

Warden is a plug-and-play moderation service. A client posts text or an image, Warden answers in under 50 ms, and isolated workers score the content in the background. Flagged items are held or removed as soon as a score lands, and the client hears about it over a WebSocket.Warden is a plug-and-play moderation service. A client posts text or an image, Warden answers in under 50 ms, and isolated workers score the content in the background. Flagged items are held or removed as soon as a score lands, and the client hears about it over a WebSocket.

It is built for small teams who want moderation as a service they run themselves, not a third-party API with per-call pricing.It is built for small teams who want moderation as a service they run themselves, not a third-party API with per-call pricing.

1. Accept first, judge later. The HTTP path never waits for inference. It validates, enqueues and returns, so a traffic spike slows the queue, not the users.

   Accept first, judge later. The HTTP path never waits for inference. It validates, enqueues and returns, so a traffic spike slows the queue, not the users.
2. Hash before you think. Every image gets a perceptual hash checked against a Redis set of banned media. Known spam is rejected in O(1) without touching the GPU.

   Hash before you think. Every image gets a perceptual hash checked against a Redis set of banned media. Known spam is rejected in O(1) without touching the GPU.
3. Two stores for two jobs. PostgreSQL holds the lifecycle of each request; MongoDB takes the noisy per-model confidence telemetry.

   Two stores for two jobs. PostgreSQL holds the lifecycle of each request; MongoDB takes the noisy per-model confidence telemetry.
4. Lazy weights. Model weights are baked into the image but loaded on the first task, which keeps idle workers small.

   Lazy weights. Model weights are baked into the image but loaded on the first task, which keeps idle workers small.

ingest response, p95ingest response, p95

<50ms<50ms

<50ms<50ms

pHash fast path for known mediapHash fast path for known media

O(1)O(1)

O(1)O(1)

models: text toxicity + NSFW visionmodels: text toxicity + NSFW vision

22

22

• The p95 figure comes from local load tests, not production traffic.

  The p95 figure comes from local load tests, not production traffic.
• Toxicity models carry the biases of their training data; borderline scores go to a human review queue rather than auto-removal.

  Toxicity models carry the biases of their training data; borderline scores go to a human review queue rather than auto-removal.
• Next: per-tenant thresholds and a small dashboard for reviewers.

  Next: per-tenant thresholds and a small dashboard for reviewers.