Content moderation for business
Content moderation that follows your rules.
Trained on your data.
Off-the-shelf filters don't know your business, so they block too much or let too much through. We tune a model on your content and your policy. Every decision - allow, review or block - comes with a confidence score and a reason.
The short version
Where it runs
A custom model is deployed in your own environment: on-prem, private cloud or an isolated VPC.
With the API option, an API provider processes the text under the scope agreed in the contract. We help you pick the option that fits your security requirements.
How we tune it
A custom model is fine-tuned on your data. With the API we set thresholds against your own targets.
Either way we account for your content types, user languages and risk level, so the output actually helps your moderators.
What you get back
A decision: ALLOW, REVIEW or BLOCK, plus a confidence score and a reason.
Not just a number to interpret. Every decision is logged, so you can explain it later and report on it.
Why off-the-shelf filters fall short
They don't know your business
A generic model has no idea what your customers, your language and your risks look like. So it rejects legitimate reviews and lets through spam that's specific to your industry.
They give you a score, not a decision
Nobody wants to interpret a 0.73. A moderator needs to know: publish, check, or block - and why.
Nobody keeps them up to date
Spam and filter evasion change every month. A model that doesn't get corrections gets worse over time, and usually no one notices.
Why isn't a general-purpose AI enough?
You can send a listing to ChatGPT and ask whether it is fine. For a handful of texts that works. At millions of texts, four problems appear.
You pay per token
General models are billed per token, so the bill grows with every word of your policy pasted into the prompt. With us you pay per item processed or a flat subscription, and the rules live in the model and the thresholds, not in the prompt.
An opinion instead of a decision
A general model replies in free text you have to parse. We return ALLOW, REVIEW or BLOCK, a confidence score and a reason in a fixed format.
You set the thresholds
You decide how many cases go to a human. Changing a threshold does not mean rewriting a prompt or retesting every answer.
An audit trail
Every decision is stored with its reason and the rule version, so you can explain it later.
A general model isn't bad. At small scale it is often enough. The difference starts when cost at high volume, consistent decisions and the ability to explain them matter. We run the comparison on your data in a pilot, not on slides.
What we do
Content moderation
For companies with comments, reviews, listings or user-submitted content. We build a system that decides what to publish, what to check by hand and what to reject.
Language model deployment
For companies that want AI for writing, editing documents or automating correspondence. We pick the model, connect it to your processes and make sure it writes in your brand's voice.
Two ways to start
A: Your own model
- Best for: companies that want full control over the model, the rules and the environment.
- Pilot: usually 4–10 weeks.
- Data: we need your data and a test set from your domain.
- Control: highest (model, rules, retention, audit).
- Cost: higher upfront, the most flexibility in the long run.
B: Ready-made API
- Best for: teams that want to start fast and check whether it pays off.
- Pilot: usually 2–6 weeks.
- Data: you don't need a full training set. A sample is enough to set thresholds.
- Control: medium (you set the rules, but you don't change the model itself).
- Cost: lower entry point, faster results.
What a decision looks like
An illustrative example. We set the rules with you, and the response has a fixed format.
Electronics
Suspected counterfeits or off-platform contact details: BLOCK or REVIEW, depending on the threshold.
Supplements
Claims of curing diseases: REVIEW, with a more cautious threshold than in other categories.
Fashion
Brand misuse in a description: REVIEW while confidence is low.
Example use cases
Social and community platforms
We filter comments and posts before other users see them. Different rules for different risk levels.
Marketplace and e-commerce
We check listings, reviews and seller content. We catch fake reviews, spam and attempts to game the catalog.
Media and publishers
We check articles and user submissions before they go live. Moderators get a reason, not just a flag, so they decide faster.
Finance, legal, compliance
For high-risk content, every decision is stored with its reason, model version and timestamp. Easier to show an auditor.
How we work
1. A conversation
You tell us what content you moderate, how decisions are made today and where it hurts. We tell you whether we can help and how. If it doesn't make sense, we'll say so.
2. A pilot on your data
We run the system on a sample of your content and measure the results. You see what works before you commit to anything longer.
3. Production and fixes
After launch, errors and borderline cases go back into the model. We keep correcting it.
Start with a pilot on your data
Bring a sample of your content, for example 1,000–10,000 listings. You see the result before you commit to anything longer.
We set the rules
What is allowed, what is not and what should go to a human.
We run it on your content
The system scores the sample and we review borderline cases together.
We measure the result
You get a report: how many decisions can be automated and where the system gets it wrong.
You decide
We roll it out, or we stop at the report.
Where your data goes
On your servers or in a private cloud
We deploy inside your environment, so you control it.
Decision history with reasons
Every decision is stored with its reason. Easy to explain later or show to an auditor.
Model and rule versions
We record which model version and which rules were active on any given day.
Retention and access
You decide how long data is kept and who can access it.
No data leaving your environment (own-model option)
With your own model, your content does not have to leave your environment.
FAQ
Which option should we pick?
If control over the model and the rules matters most, go with your own model. If you want to see quickly whether it works for you, start with the API.
Do we need our own data?
With the API, no. A sample of your content is enough to set thresholds. With your own model you do need data, because quality depends on it.
Does our data have to go to a public cloud?
No. After go-live the system can run entirely in your environment. We'd only use a public cloud during the pilot, and only if you want that.
How long does a pilot take?
Usually 2–6 weeks for the API and 4–10 weeks for a custom model. It depends on data quality, whether you have labelled examples, and how much integration is involved.
Can we start with the API and move to our own model later?
Yes. The response format stays the same (ALLOW/REVIEW/BLOCK), so you don't have to rebuild the integration.
Do I pay per token?
No. We charge per item processed or as a subscription, with no token metering. Cost doesn't grow with the length of your rules. We quote a specific rate in the proposal.
Is your own model cheaper than a large-model API?
It depends on volume and text length. At small scale a general model can be cheaper. That is why we compare costs on your data during the pilot.
For specific use cases
Pages on the most common ways our customers use it.
- Content moderation API - An ALLOW, REVIEW or BLOCK decision for every text. Plugs into the process you already have.
- Marketplace and e-commerce moderation - Listings, reviews and seller content: less spam and manipulation in your catalog.
- Medical and healthtech content moderation - Health content needs more cautious rules and a record of every decision.
- Review and seller content moderation - Buyer reviews and seller messages: protect credibility without blocking honest criticism.
- How a moderation pilot works - You bring a content sample, we measure the result. What a pilot includes, how long it takes and what you get.
- Moderation policy and thresholds: ALLOW, REVIEW, BLOCK - What each decision means, how confidence thresholds work and how to fit them to your team capacity.
- Own model or an LLM with a prompt? A guide - When a large model with a prompt is enough for moderation and when a model tuned on your own data pays off.
Articles
Short pieces on rolling out content moderation and setting ALLOW/REVIEW/BLOCK decisions.
- Why companies need a content moderation API - Comments, reviews and tickets are increasingly written by language models. You can't read them all by hand.
- Content moderation: when manual review stops being enough - Most companies start with moderators and a list of banned words. It works until the platform grows.
- What “training a model on your data” actually means - You don't upload a CSV and press train. First we clean the data, then we train, then we set the thresholds.
Got a content problem? Describe it.
Tell us what isn't working. We'll reply with a concrete proposal, not a sales pitch.