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
Seen this comment before? Instant verdict from cache. Zero model cost. Zero latency.
Classifies the comment as English, Hinglish, Punjabi, or Bangla — or flags it as uncertain. Gets the language right before any moderation model sees it.
English → MiniLM-lite. Hinglish → MiniLM-lite + HinTox verification. Punjabi / Bangla → dedicated local classifiers. Each with independently calibrated confidence thresholds.
Model disagreement or low confidence? The comment goes to a large language model for a second opinion — or to human review. Never auto-approved.
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 moderation | Keyword blocklists | CleanFeed India | |
|---|---|---|---|
| Detects Hinglish abuse | Partial (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 detection | ❌ | If manually added | ⚠️ Under development |
| Cost per comment | Free but ineffective | Free but incomplete | Cache-first = near-zero for repeats |
| False negatives → | Abuse stays live | Abuse stays live | Goes to review, never auto-kept |
Design principles
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.
Cache-first economics. Repeat comments return instant verdicts from a versioned cache. No model call needed. This makes blocking serial abuse effectively free.
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.
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.
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.
We’ll reach out when a spot opens.