Instagram’s moderation was built for English. Your comment section wasn’t.

CleanFeed India is an AI comment moderation tool that understands abuse in the languages Indian creators actually type in: English, Hinglish, Romanized Punjabi, and Romanized Bengali.

It doesn’t adapt an English model and hope for the best. It routes every comment to the right language-specific model — then hides the hate, keeps the conversation, and never silently approves something dangerous.

How it works

The moderation pipeline

1
Cache check
Seen this comment before? Instant verdict from cache. Zero model cost. Zero latency.
2
4-language FastText router
Classifies the comment as English, Hinglish, Punjabi, or Bangla — or flags it as uncertain. Gets the language right before any moderation model sees it.
3
Language-specific moderation
English → MiniLM-lite. Hinglish → MiniLM-lite + HinTox verification. Punjabi / Bangla → dedicated local classifiers. Each with independently calibrated confidence thresholds.
4
Escalation (only when needed)
Model disagreement or low confidence? The comment goes to a large language model for a second opinion — or to human review. Never auto-approved.
5
Action
Hide. Keep. Or flag for review. The verdict and reasoning are stored in cache so repeat abuse is blocked instantly next time.

What makes this different

Meta’s moderationKeyword blocklistsCleanFeed India
Detects Hinglish abusePartial (native Hindi only)Only if exact word is listed✅ Purpose-built
Detects Romanized Punjabi/Bangla✅ FastText router + local classifiers
Handles abbreviations (“bc”, “mc”)If manually added✅ Model-based, not keyword-based
Handles emoji substitution✅ Learned patterns
Caste-based abuse detectionIf manually added⚠️ Under development
Cost per commentFree but ineffectiveFree but incompleteCache-first = near-zero for repeats
False negatives →Abuse stays liveAbuse stays liveGoes to review, never auto-kept

Design principles

  1. Language routing before moderation. The biggest source of false negatives in Indian comment sections is processing Hinglish with an English model. We fix the language first.

  2. Cache-first economics. Repeat comments return instant verdicts from a versioned cache. No model call needed. This makes blocking serial abuse effectively free.

  3. Never silent-keep. If every model is uncertain, the comment goes to human review — never auto-approved. In child safety, a false negative is always worse than a false positive.

  4. Calibrated confidence, not raw scores. A softmax output of 0.7 is not a 70% probability. We calibrate thresholds per model per language on held-out data before any auto-hide is enabled.

  5. No model failure = keep. If a model crashes, times out, or returns invalid output, the comment is queued for review. It is never silently approved.

Status: Private beta

CleanFeed India is currently in private beta. We’re onboarding a small group of Indian creators to test the moderation pipeline on real comment sections.

Join the waitlist →

We’ll reach out when a spot opens.