Argus: a comment moderator that tells you what you could say
Most moderators say no. Argus says why, and shows you what you could say instead. Built on NVIDIA Nemotron.
Argus is a drop-in comment moderator. When someone writes a rude comment, it doesn't just block it. It shows a better way to say the same thing, and the writer chooses Use this, Edit or Ask a moderator. When a comment is genuinely harmful (a threat, hate, someone's private phone number), it isn't posted, and the writer is told what's wrong and why so they can rewrite it themselves.
Every AI decision runs on NVIDIA Nemotron 3 Ultra, served by Nebius Token Factory.
- Hackathon: Nebius x NVIDIA Global AI Hackathon 2026, Best Apps and Agents track
- Three parts, one engine:
plugin/ is the drop-in moderator for websites (engine + server handler + widget, zero dependencies); extension/ is a browser extension that blocks the Post button on YouTube, Reddit and Instagram (and, if you turn it on, WhatsApp Web) until a comment is allowed; and the Next.js app is a demo site plus the moderator queue
What the writer sees
| They write | What happens |
|---|
| "Great news, congrats!" | Posts |
| "i hate you" (on a post about singing) | Not posted yet. This tells another person you hate them. Suggestion: "I'm not a fan of your singing." → Use this / Edit / Ask a moderator |
| "Women shouldn't be allowed to sing on stage. Go back to the kitchen." | Can't be posted. This says women can't perform because of their gender. Singing ability has nothing to do with gender, and comments like this push women out of music. → Edit my comment (no AI rewrite; also sent to a moderator) |
| "enna idhu da" (Tanglish the AI can't read with confidence) | Not posted yet. We couldn't be sure what this means, so a person will check it. → a moderator approves or keeps it unposted |
| (the AI is unreachable) | Not posted. "We couldn't check this comment right now." It fails closed, never open |
How it works
comment → Nemotron scores the words alone (0..1 + category + plain-language reason)
├── clean (< 0.40) → posts
├── rude (0.40–0.85) → reason + a checked rewrite: Use this / Edit / Ask a moderator
├── severe (≥ 0.85) → explanation of what's wrong and why; writer edits it; moderators notified
├── unsure → a non-English comment the AI can't read with confidence: a moderator decides
└── model down → nothing posted, writer told to try again
Ask a moderator / unsure → waits unposted; no answer in 10 minutes → stays unposted (fails closed)
A suggestion is only shown if it passes three checks: it's in the same language and script as the original (Hindi in, Hindi out), a second Nemotron call confirms it keeps the writer's meaning, and it scores clean itself. The suggestion can use the post being replied to ("I'm not a fan of your singing"), but the verdict never does: every comment is judged on its own words.
Other safeguards: emails and phone numbers are masked before text reaches the model; an instant keyword check flags obvious threats while typing and can only make a verdict stricter; the plugin's endpoints have an optional secret key and a per-IP rate limit, so nobody can spend your model credits.
Using NVIDIA Nemotron on Nebius Token Factory
All model calls go through plugin/src/nemotron.mjs, an OpenAI-compatible client pointed at Token Factory with nvidia/Nemotron-3-Ultra-550b-a55b, reasoning off, temperature 0, with retries on rate limits. Three prompts in plugin/src/prompts.mjs: scoring (including the "explain why" text for severe comments, so that costs no extra call), rewrite, and the meaning check.
Evaluation
scripts/eval-set.ts holds 85 hand-labelled comments: praise, disagreement, profanity used for emphasis, reports of abuse, insults, sarcasm, violent hyperbole, prompt-injection attempts, threats, doxxing, self-harm encouragement, Spanish, and Indian languages — Tamil in its own script and typed in English letters (Tanglish, including compliments vs insults and alternate spellings), Hindi and Hinglish.
Current (85 cases, Nemotron 3 Ultra): 98.8% (84/85), severe caught 17/17, 0 failed closed. Two Tanglish comments the model couldn't read with confidence went to a person instead of being guessed.
The model comparison below was run on the first 63 cases, fresh runs, no cache:
| Model on Token Factory | Accuracy | Severe caught | Failed closed |
|---|
| Nemotron 3 Ultra (550B-A55B) ✅ used | 98.4% (62/63) | 15/15 | 0 |
| Nemotron 3.5 Super-VL | 98.4% | 15/15 | 0 |
| Nemotron 3 Super (120B-A12B) | 95.2% | 15/15 | 0 |
| Nemotron 3 Nano (30B-A3B) | 88.9% | 15/15 | 0 |
| Nemotron 3.5 Lightning | 87.3% | 14/15 | 2 |
| MiniMax M3 (original prototype model) | 98.4% on its own API, 93.7% via Token Factory | 15/15 | 0 / 4 |
The one miss ("You're a pathetic excuse for a developer, delete your account") scores exactly on the severe line, so it errs toward explaining rather than suggesting. Turning Nemotron's reasoning off mattered: with it on, the same engine scored 82.5%, because the model sometimes answered the comment instead of judging it.
Run it
You need Node 20+, a Nebius Token Factory API key and, for the demo app, a Supabase project.
Just the plugin (no database):
cd plugin && npm test
cd plugin && NEBIUS_API_KEY=your-key npm run demo
Open http://localhost:8790 for the drop-in widget on its own. See plugin/README.md to add it to your own site.
The demo app:
npm install- Copy
.env.example to .env.local and fill in the Nebius key and the three Supabase values.
- In the Supabase SQL editor, run the files in
supabase/migrations/ in order.
npm run seed creates Maya and Alex, three posts and a few replies.npm run dev, then open http://localhost:3000. The moderator queue is at /mod.
| Command | What it does |
|---|
npm test | Plugin tests with a stub model (no network, no cost) |
npm run eval | The 85-case set against the real model |
npm run seed | Reset the demo data |
Project layout
plugin/ the product: engine, Nemotron client, prompts, HTTP handler, widget, tests, demo
app/ demo app pages (feed, /mod) and API routes
extension/ Chrome extension (YouTube, Reddit, Instagram, WhatsApp Web) + test bench
components/ composer + suggestion card, comments, review cards
lib/ config, the app's moderator instance, comments, review queue, feed
scripts/ seed, eval, eval set
supabase/migrations/ schema
Built during the hackathon
The project began on Aug 23, 2026, just before the submission period opened on Aug 26. Built or significantly updated during the submission period:
- Moved every model call to NVIDIA Nemotron 3 Ultra on Nebius Token Factory, after evaluating all five Nemotron models available there
- Rebuilt the moderator as a drop-in plugin with zero dependencies (engine, server handler, widget, tests)
- Severe comments are now explained (what's wrong and why) instead of hard-blocked, with no AI rewrite
- Suggestions use the post being replied to, and are checked for language, meaning and safety before they're shown
- Tamil, Tanglish, Hindi and Hinglish: read correctly, suggestions in the writer's own language, and comments the AI can't read with confidence go to a moderator instead of a guess
- The browser extension for YouTube, Reddit, Instagram and WhatsApp Web
- Fail-closed behaviour, privacy masking, a rate-limited and optionally secret-protected API, and a moderator queue with timeouts (in the demo app and in the plugin itself)
Nebius and NVIDIA tools used
- NVIDIA Nemotron 3 Ultra (
nvidia/Nemotron-3-Ultra-550b-a55b): every AI call (scoring, rewrite, meaning check)
- Nebius Token Factory: serves the model through its OpenAI-compatible API. Switching models was a one-line change, which is how we ran the full eval against five Nemotron models and picked on evidence
- No Nebius AI Cloud
Feedback
- Token Factory's OpenAI-compatible API made switching models a one-line change, so comparing five Nemotron models on our 85-case eval took an afternoon.
- Nemotron 3 Ultra was the most accurate and wrote the best rewrites. Nano was fast but weaker on sarcasm and Indian languages.
- Reasoning off mattered a lot: with it on, accuracy dropped to 82.5%, because the model sometimes replied to the comment instead of judging it. A note about this in the docs for classification tasks would help.
- Indian languages typed in English letters (Tanglish) needed explicit examples in the prompt; without them some compliments were read as insults.
- We hit occasional rate limits and server errors and added retries; published per-tier rate limits would help.
- The same third-party model scored lower through Token Factory than on its own API (MiniMax M3: 93.7% vs 98.4%), which may be worth a look.
License
MIT
