We build the systems that keep your platform safe.

1Intake

Every post, message, image and profile enters one queue, with user and context attached.

Every item, one place

• Posts, messages, images, profiles and reports
• User age, history and trust signals attached
• One schema across every product surface

2Fast filters

Hash matching for known abuse, plus scam and spam rules. Most traffic is settled here in milliseconds.

p95 under 20 ms

• Perceptual hash matching (PDQ, PhotoDNA)
• Scam, link, phone and payment handle rules
• Burst and rate limits on new accounts

3Classifiers

Models tuned on your data for text, images and behavior. They score risk, they do not decide alone.

Tuned per policy

• Text, image and behavior models
• Vendor APIs or models trained on your labels
• Thresholds tuned per harm, with a review band

4Policy judge

An LLM reads your written policy and decides the hard cases. Every decision comes with a reason.

A reason for every call

• Your written policy is the prompt
• Verdict, category and a cited reason
• Low confidence always goes to a person

Example reason: Asks to move off platform after 2 messages. Policy 4.2, off platform solicitation.

5Human review

Only truly ambiguous items reach your team, ranked by risk, with the context already gathered.

Reviewers see what matters

• Queue ranked by risk and reach
• Context and history in one view
• Blurred previews and exposure limits

6Action and reporting

Takedowns, appeals, NCMEC reports and 48 hour removal clocks, logged for auditors.

Audit trail by default

• Remove, restrict, warn, ban or escalate
• 48 hour TAKE IT DOWN removal clocks
• NCMEC reports, appeals and audit logs

7Evals and cost

Every decision measured for accuracy, speed and cost, so the system gets cheaper and better each month.

Cost per decision, tracked

• Golden sets labeled with your team
• Precision and recall per harm
• Cost per decision for every stage