Paste Any Spam on Our Landing Page. We'll Classify It Live.

Every moderation-bot landing page says the same things. "99% accuracy." "AI-powered." "Context-aware." You read it, you believe some of it, and then you still don't know if the bot will survive contact with your actual group.

The reason is simple: you can't test a moderation bot properly without installing it. And installing it means giving admin permissions to something you haven't proven. It's a trust gap baked into how these tools ship. Every admin I've talked to has the same story — they try three bots over six months, each of them over-bans real members once, and they go back to moderating by hand.

So we did something nobody else does. We put our actual production AI classifier on our landing page. Not a recorded demo. Not three hand-picked examples. The same model that's running right now in 41 live Telegram groups, catching ~180 spam messages per week. You paste any message, and you get the real verdict in 3 seconds.

Try it — paste whatever you want

Right now, below this paragraph, there's an input box. Paste something. Anything. A scam DM you've received. A message a keyword bot wrongly deleted from your group. A normal conversation. A borderline case that's hard to judge. The AI will tell you what it'd do — delete, ban, delete-only, ask the admin, or leave it alone — and explain why.

Go ahead. We'll wait. Then come back and read the rest.

Why this is unusual

Almost every B2B tool with an AI component ships a "demo" that's actually a pre-computed fake. You click "Try it free," you get a modal with three canned examples, you click "Run analysis," and it plays back a result that was baked in HTML. It looks like AI, but it's just marketing.

We understand why people do this. Exposing a real AI endpoint publicly is scary:

• Cost. Every paste costs a fraction of a cent in LLM calls. At scale, that's real money.
• Abuse. Someone can turn your demo into a free classification API — run 10,000 requests and burn your budget.
• Embarrassment. The model sometimes gets things wrong. In production, you log the error and iterate. On a public demo, every wrong answer is a screenshot someone can tweet.

So most companies ship the fake. We shipped the real one. Here's how we handled each concern:

Cost: per-IP rate limit (3 classifications per hour), request size caps (10 KB text, 2 MB image), and a daily budget ceiling with alerting. Worst case we've modeled is ~$20/month even at 100 unique testers per day. If Varta converts one additional customer per week because the demo closed them, the demo pays for itself ten times over.

Abuse: origin allowlist (requests only accepted from getvarta.com), Cloudflare Turnstile ready to enable the moment we see anything weird, and the rate limit is strict enough that scraping the endpoint for free classification is slower than hosting your own model.

Embarrassment: we accept it. If the classifier gets something wrong in front of a prospect, that's data. The alternative — hiding behind a canned demo — means we'd never know what edge cases trip us up until a paying customer tells us at 2 AM.

What's actually happening under the hood

The classifier runs on Claude Sonnet (primary) with GPT-4.1 and Gemini as fallbacks. Same LLM stack that handles real moderation decisions in production. The system prompt asks for structured JSON: a verdict (spam / clean / uncertain), an action (what the bot would do — delete, ban, delete-only, ask admin, or none), a category, 2-4 plain-language reasons, and one "teaching moment" explaining why a keyword-only bot would likely get the answer wrong.

If you forward a message from another group instead of pasting text, the classifier also gets the original sender's username. We cross-check that username against our ban database across all 41 groups. If this author has been banned in 2+ of our groups in the last 30 days, you'll see an extra bullet: "I know this author — banned in X of my groups this month." That's signal no other moderation bot can give you, because no other moderation bot operates across a shared cross-group intelligence network.

For screenshots, we use a vision model that can parse multiple messages from a single image — top-of-chat context, a spam message in the middle, legitimate replies below — and give you a verdict on each one separately.

Why there's no confidence percentage

Most AI demos show "87% confidence" next to the verdict. We tried that at first. Then we watched five non-technical admins read it, and every single one got confused. "So it might delete it? It might not? What does 87% mean — the model's 87% sure, or it'll delete 87% of the time?"

So we threw it out. Instead, the verdict maps directly to what the bot would actually do in your group, in plain language:

• 🚨 SPAM — I'd delete and ban. Confidence high enough to trigger autonomous mode.
• 🗑 Spam — I'd delete. Delete-only mode; wait for third strike before banning.
• 🤔 Borderline — I'd ask you. Cautious mode; admin judgment needed.
• ✅ Clean — I'd let it through. No action.

These map one-to-one to the four moderation modes you can set in /settings after adding Varta. Seeing the verdict is seeing what would happen. No translation from probability to action required.

What you should actually test

Most people paste obvious spam. It gets flagged. Cool. But the interesting tests are:

False positives from keyword bots. Paste a message a keyword bot wrongly banned. "I saw free pizza at the office kitchen." "Guaranteed no-cost seminar next Tuesday." Real messages that contain spam-trigger words. A context-aware AI should pass these. If ours flags them wrong, tell us — we'll use your paste to improve the model.

Borderline ads. Someone in your dev community posts "I'm hiring a senior Go engineer, remote, $120k." Is that spam? Depends on whether hiring is allowed in your group. Paste it, see what the AI calls it. You'll probably get "borderline — ask admin," which is the honest answer.

Your specific language. Ukrainian. Turkish. Arabic. Hindi. The model isn't trained on "Telegram spam patterns" — it's trained on language, and it reasons about intent. Paste something in your language and watch the reasons come back in your language too.

Actual scams you've screenshot. Upload the screenshot (coming soon — or forward the message via the bot's DM for now). The vision model will parse out every visible message and tell you what it'd do with each.

What shipping this publicly signals

Two things, really.

First — we think the model is good enough that watching strangers test it isn't embarrassing. When we ran the numbers internally, we were seeing 94-97% accuracy across the typical spam categories. Not 99%, but calibrated. The model knows when it doesn't know.

Second — we believe the best thing a moderation tool can do is be testable before it's trusted. Installing a bot into a live group and hoping it doesn't nuke your best customer is a terrible trust model. Paste-to-classify replaces it with: test the model first, then install only if it impressed you.

If you paste 5 messages and the verdicts are bad, don't install Varta. Tell us what we got wrong instead, and install something else. We'd rather you decide based on evidence than based on a landing-page claim we made about our own accuracy.

One more thing

When you click "Add Varta to your group" after seeing a verdict, the link includes a session hash. That hash lets us tie "this anonymous paste" to "this bot install" to "this first spam catch in their group" — a full conversion funnel we can measure and improve. It's the same philosophy: we'd rather measure reality than guess.

If you paste something and the verdict surprises you (good or bad), we'd love to hear about it. Message me directly on Telegram — @darynakulya. Real feedback from real admins is how this gets better.

Varta is a Telegram moderation bot that learns your group's rules and runs the same AI live — on this page, in your group, everywhere. Add Varta for free →