EVOMIND
Log in
Capabilities

Every moderation tool you need, one endpoint

Five models, each fine-tuned for one slice of moderation. Same request format, same billing, same rules-first pipeline.

5
task-specific models
1
API endpoint
256
token context window
<200ms
typical latency
Pipeline

How a request is processed

Input

Your text is sent to POST /v1/moderate with the jailbreak model.

Click a stage to see what happens at that step. Switch the model above to see how the rules and transformer handle each type of content.

Side by side

How the five capabilities differ

AttributeJailbreakVulgarPIIScamHarmful
What it catchesPrompt injection and instruction overridesProfanity and targeted abusePersonal information in textPhishing, fraud, and impersonationInstructions for real-world harm
Where it mattersLLM chat surfaces and copilotsPublic comments and live chatAssistants handling customer dataMessaging and marketplacesBroad or young audiences
Rules filter catchesKnown injection phrasingExact profanityEmails, cards, ID formatsCommon scam phrasingExplicit harm phrasing
Transformer resolvesObfuscated and encoded attemptsContext and targeted insultsContextual data requestsNuanced scam structuresSupport-seeking vs harm
Typical latency<200ms<200ms<200ms<200ms<200ms

All five share one endpoint, one billing rate, and the rules-first pipeline. Only the task they are fine-tuned for changes.

Deep dive

The technical details, model by model

jailbreak

Jailbreak detection

Catches attempts to override a system's instructions, extract hidden prompts, or push an assistant into ignoring its guardrails, whether stated directly or disguised through roleplay, encoding, or indirect phrasing.

Catches
  • Direct instruction overrides such as "ignore all previous instructions"
  • Roleplay jailbreaks and system prompt extraction
  • Obfuscated or encoded override attempts
Technical
  • Model ID: jailbreak in POST /v1/moderate
  • Rules filter runs first, then a task fine-tuned transformer
  • 256-token window with overlapping chunking for longer inputs
vulgar

Vulgar language

Detects profanity from mild to severe, along with targeted insults and harassment aimed at a specific person. Genuine discussion or education about sensitive language still passes.

Catches
  • Mild to severe profanity
  • Targeted insults and harassment directed at a person
Technical
  • Model ID: vulgar in POST /v1/moderate
  • Rules filter runs first, then a task fine-tuned transformer
  • 256-token window with overlapping chunking for longer inputs
pii

PII detection

Identifies personal information such as emails, phone numbers, government IDs, card numbers, and addresses, whether a user is sharing their own or asking for someone else's.

Catches
  • Emails, phone numbers, and physical addresses
  • Government IDs, credit cards, and bank details
  • Requests asking for someone else's personal information
Technical
  • Model ID: pii in POST /v1/moderate
  • Rules filter runs first, then a task fine-tuned transformer
  • 256-token window with overlapping chunking for longer inputs
scam

Scam detection

Recognizes common scam structures: phishing, advance fee fraud, fake investment pitches, romance scams, and impersonation of trusted institutions. Genuine questions about spotting scams still pass.

Catches
  • Phishing attempts and fake login pages
  • Advance fee, lottery, and crypto scams
  • Impersonation of banks, government, or companies
Technical
  • Model ID: scam in POST /v1/moderate
  • Rules filter runs first, then a task fine-tuned transformer
  • 256-token window with overlapping chunking for longer inputs
harmful

Harmful content

Flags instructions or facilitation of real-world harm, including weapons, violence, extremism, and self harm, while carefully preserving genuine support-seeking language.

Catches
  • Instructions for weapons or explosives
  • Violence planning or facilitation
  • Self harm facilitation, not support-seeking language
Technical
  • Model ID: harmful in POST /v1/moderate
  • Rules filter runs first, then a task fine-tuned transformer
  • 256-token window with overlapping chunking for longer inputs

Try all five for free

Sign up, get 1,000 free requests, and run every model against real text before you decide anything.