Solutions / Content moderation

Content moderation that follows your rules, not someone else's.

Every platform draws the line somewhere different. Nyckel learns where you draw it — from the calls your team already makes — and applies it to every image and every message that arrives.

Get started free Talk to us
Taimi logo Gust logo RetailHub logo PMG logo Pulse For Good logo

Why generic filters don't work

Same subject. Opposite answers.

Both of these are photos of a dog. One passes and one fails — and nothing about the subject explains why.

Nyckel console: a puppy photo passes at 99.8% confidence, a dog photo with emoji stickers over it fails at 100%
What separates them is the sticker overlay — and whether that's allowed is a policy question only you can answer. An off-the-shelf NSFW filter has no opinion about it at all. A function trained on your own pass and fail examples does, because your examples are the policy.

This is the whole argument for custom moderation. The obvious violations are the easy part. Your rules live in the cases a general-purpose model was never built to judge.

Images and text

The same thing works on what people write.

Nyckel console classifying two messages: one as Spam at 91.6% confidence, one as a Promising Lead at 94.6%
Spam, scams, harassment, off-topic posts, or your own categories entirely — text functions return a label and a confidence score exactly the way image functions do. A listing with a photo and a description can be judged as one item using a tabular function.

Why it's worth automating

What makes this hard by hand.

The rules are yours
What's fine on one platform gets someone banned on another. A general-purpose moderation API encodes somebody else's policy, not yours.
The hard cases are the point
Nobody needs help with the obvious violations. The value is in the ambiguous middle — and that's exactly the part a written policy captures worst.
Volume makes it expensive
Reviewing everything by hand doesn't scale. Reviewing nothing isn't an option either.
The line keeps moving
New tactics, new slang, new ways around the filter. A static model goes stale; one that retrains on your reviewers' corrections doesn't.

How it works

Three steps.

1

Show it your calls

Label a few examples the way your team would. The corrections your moderators already make every day are training data — you're not creating new work, you're capturing work that currently evaporates.

2

Set a confidence threshold

Auto-clear the confident approvals, auto-flag the confident violations, and send only the uncertain middle to a person. See tuning thresholds for how to pick the numbers.

3

Watch the queue shrink

Every reviewed item feeds the next retrain, so the band of cases needing human eyes narrows over time. Your reviewers spend their attention where it actually matters.

In production

Taimi moderates images and text on Nyckel.

A dating platform with user-generated photos and messages — the case where getting moderation wrong is most visible, in both directions.

Read the Taimi case study → Try a ready-made moderation model on your own data → Build your own classifier → Wire it up — API quickstart →