Automated content moderation for websites

Relevio

The context‑aware, rule‑driven semantic moderation engine for autonomous story boards and forums

Relevio is a unique context‑aware, rule‑driven moderation engine that enables fully autonomous or semi-autonomous forums, story boards and social networks.

With Relevio you can reduce moderation workload dramatically, eliminate your moderators exposure to distressing content and stamp out coordinated / agentic spam.

Take away the burden of moderatating your forum today.

Products

How does it work?

Relevio is a new class of moderation engine employing Rule‑Driven Semantic Moderation, which uses the full power of large language models (LLM). This is very different from the established method of Category‑Based Classifier Moderation employed by systems such as Hive or Tisane.

Uniquely, Relevio is invoked from the user interface before content ever reaches your server. Your server only ever receives quality checked, safe content.

By preventing bad posts from being created in the first place, Relevio reduces your workload in absolute terms and it protects your moderation team from being exposed to harmful or distressing content.

By contrast CBCM moderators can only work on content after it reaches your server. By automatically flagging bad content or moving it to your moderation queue, such systems can actually increase your moderation workload. Moreover, the expectation of human intervention means workers are still exposed to harmful content.

Relvio can also work in pipe-lines or on historical data to tackle your pre-existing moderator workloads.

You define the rules in plain English. Relevio applies them consistently, using AI only as a semantic interpreter. There is no drift, no unpredictability and no black‑box behaviour. There are no fixed categories and no percentage scores. Each rule either passes or fails.

How does Relevio stack up?

Here's how Relevio compares with Category‑Based Classifier Moderation.

Dimension	Relevio	Category‑Based Classifier
Where moderation happens	Invoked from the browser — content is checked before submission	Server-side — content is ingested first, then classified
Security model	Pre-submission gatekeeping — your backend only receives approved, signed content	Post-submission filtering — harmful content still reaches your server
Backend engineering	Minimal — backend only verifies a fast HMAC; no queues, pipelines, or moderator UI	Heavy — requires queues, rate-limiting, pipelines, dashboards, and human oversight
Latency distribution	Latency sits in the client, keeping your backend fast and predictable	Latency sits in the server, requiring scaling and back-pressure handling
Context awareness	Yes — story, comment, and reply context included in every check	No — evaluates each message in isolation
Image + story consistency	Multimodal semantic checking — verifies that images match the story and flags contradictions	Independent classifiers — image and text are evaluated separately with no cross-consistency checks
Decision logic	Rule-driven — deterministic pass/fail based on your ruleset	Model-driven — probabilistic scores based on opaque model weights
Output format	Structured JSON with pass/fail, error description, and suggestion	Structured JSON with Category scores (e.g., toxicity %, spam %)
Explanation quality	Rule-specific, contextual explanations that tell users exactly which rule failed and how to fix it	Category-based explanations tied to fixed labels (e.g., bullying, hate); cannot reference custom rules or full context
Input size limit	Bounded by the LLM context window — can handle long stories and rich context	Max ~1024 characters for text moderation, limiting story length and context
Cryptographic trust	HMAC-signed decisions binding content, images, ruleset, and context	None — no verifiable proof of moderation integrity
Injection resistance	Strong — Malformed rulesets, and instruction attempts are rejected. Adversarial phrasing is readily detected by the LLM	Variable - Vulnerable to adversarial phrasing — classifiers can be bypassed with obfuscation and semantic tricks
Relevance enforcement	Built-in — rules can require replies to relate to the story or parent comment	Not possible — no concept of conversational structure
User experience	Instant, transparent feedback with explanations and suggestions that reduce frustration	Category‑based feedback only. Reasons are granular and cannot reference context or provide tailored guidance
Autonomous operation	Yes — ideal for self-moderating story boards and forums	No — requires server-side moderation infrastructure