DEV Community: PersonymAi

Transparent Moderation: We Now Show Why We Ban

PersonymAi — Sat, 11 Apr 2026 16:39:40 +0000

Today we rolled out a significant improvement to PersonymAi Moderator.
Every time a user is banned, the system now displays a detailed banner containing:
• The user who was banned
• The exact reason for the ban
• Spam Score (0–100%)
The message itself automatically disappears after 60 seconds to keep the chat clean, while the full information remains visible to admins.
This gives administrators complete clarity into the moderation logic without compromising chat cleanliness.
No more guessing. No black boxes. Just transparent, explainable moderation.

Transparent Moderation: We Now Show Why We Ban

PersonymAi — Sat, 11 Apr 2026 16:33:53 +0000

How We Detect Reaction Spam in Telegram Using Behavioral Scoring

PersonymAi — Tue, 07 Apr 2026 14:19:41 +0000

Most Telegram anti-spam bots are built around one assumption: spammers
write messages. So we match text against patterns, run it through NLP,
check for suspicious links. But what happens when the spammer sends no
text at all?

Reaction spam is exactly that. A bot joins your group silently, then
floods every post with 🤡, 18+, and gambling emojis — harming your
channel's reputation without triggering a single keyword filter.

Our approach at ModerAI: instead of analyzing message content, we score
behavioral signals. Things like — does this user react but never comment?
Are they reacting from outside the group via channel post comments? Does
their bio contain obfuscated text patterns? Each signal contributes to a
spam probability score. Cross a threshold — you get restricted.
No text needed.

What behavioral signals have you found most reliable for detecting
non-text spam in group chats?

Building a Product You Can Never Demo Publicly

PersonymAi — Mon, 06 Apr 2026 17:49:37 +0000

What happens when your product's core value proposition requires absolute secrecy about who uses it?

At PersonymAI, we built an AI system that generates natural Telegram comments using 1,000+ unique personas — each with distinct writing styles, opinions, and behavioral patterns. The result? Neither marketers nor advertisers can tell it apart from real conversation.

Our clients use this to maintain active-looking communities and sell advertising. But this creates a fundamental marketing paradox: showing a real client channel as a case study would immediately undermine the product's value for that client.

We chose NDA over growth metrics. Every client is protected. No public channels, no named testimonials, no before/after reveals.

How do you market a product that works best when nobody knows it exists? Curious to hear how others in the AI space handle similar transparency trade-offs.

Implementing 3-Tier Moderation for Telegram Bots

PersonymAi — Tue, 31 Mar 2026 14:23:04 +0000

Binary spam detection (spam or not spam) breaks down in active communities. A forwarded giveaway could be spam or a legitimate user sharing excitement. A message saying "write me" could be a scam CTA or an angry user. We rebuilt our Telegram moderation pipeline into three action tiers using AI for intent classification: tier 1 (ban) for clear spam with profit intent, tier 2 (mute + admin buttons) for ambiguous cases, and tier 3 (3-strike warnings) for links and forwards. The system also auto-detects chat language from linked channel posts using character-set heuristics (їєґ → Ukrainian, ыэъ → Russian). How do you handle the gray area between spam and legitimate messages in your moderation systems?

Connecting a Context-Aware Telegram Moderation Bot in 5 Steps

PersonymAi — Mon, 30 Mar 2026 10:00:11 +0000

Most Telegram moderation bots run on keyword blocklists — easy to bypass, high false-positive rate, zero context awareness. ModerAI takes a different approach: you describe your group's topic in plain text, and the NLP pipeline uses that as a context window when classifying messages. The 15-layer AI stack handles the rest — no rule-building, no regex. How does a natural-language topic description feed into a spam classification pipeline at scale?

Building a Fairer Anti-Spam System: How We Handle Links, Warnings, and New Chats

PersonymAi — Sat, 28 Mar 2026 10:54:40 +0000

just shipped three changes to our Telegram anti-spam bot (ModerAI) that fundamentally change how we handle edge cases. Here's what we built and why.

The Problem With Binary Decisions
Most anti-spam bots make binary decisions: spam or not spam. Ban or allow.

This creates two failure modes:

False positives — legitimate users banned for having a link in their bio
False negatives — spammers who learn the rules and work around them
We needed a middle ground.

Change 1: Contextual Bio Link Analysis

Before:
if "t.me/" in user.bio:
ban(user) # crude but effective... and unfair

After:
link_target = analyze_link_context(user.bio)
if link_target.category in ["spam_channel", "scam", "adult"]:
ban(user)
elif link_target.category in ["game_referral", "personal_channel", "community"]:
allow(user) # legitimate use case

AI analyzes what the link actually points to. A Hamster Kombat referral? Fine. A channel selling "guaranteed 500% returns"? Ban.

Change 2: Progressive Warning System
Instead of ban-on-first-offense, we implemented a 3-strike system:

Strike 1: delete message + warn ("у вас ещё 2 попытки")
Strike 2: delete message + warn ("у вас ещё 1 попытка")
Strike 3: ban

Exception: edited message → instant ban (no strikes)

The edit detection is key. Spammers who post "Hello everyone!" then edit to a scam link 5 minutes later get zero warnings. This pattern is always intentional.

Change 3: Fresh Chat Grace Period
When ModerAI connects to a new chat, it has zero context. Every user is "unknown."

Aggressive bio scoring on day 1 would ban half the existing members. So we added a 48-hour grace period:

chat_age = now() - chat.connected_at

if chat_age < 48_hours:
# Relaxed mode: skip suspicious bio scoring
# Still ban critical threats (adult, drugs, obvious scam)
if threat_level == "critical":
ban(user)
else:
allow(user) # gather data first
else:
# Normal mode: full scoring pipeline
run_full_analysis(user)
After 48 hours, the bot has enough context to make accurate decisions.

Results
These changes reduced false positive rate from ~0.3% to ~0.1% while maintaining 99.7% spam detection.

The key insight: fairness and accuracy aren't opposites. A system that gives legitimate users the benefit of the doubt can still be ruthless with actual spammers — you just need smarter decision-making, not stricter rules.

ModerAI: $9/month per chat. 7-day free trial.

→ personym-ai.com/moderator-ai

Questions about the implementation? Happy to discuss in the comments.

How We Built Voice and Image Spam Detection for Telegram (Technical Deep Dive)

PersonymAi — Fri, 27 Mar 2026 12:59:39 +0000

Last month we shipped three features that no other Telegram anti-spam bot has: voice message analysis, image spam detection, and anti-masking intelligence. Here's how we built them.

The Problem
Spammers evolved. Our text-based pipeline was catching 99.7% of text spam. So spammers stopped using text.

Voice messages with gambling/scam ads
Images with overlaid promotional text
Text with emoji inserted between every character
Traditional keyword and even AI text analysis is blind to all three.

Voice Message Pipeline
Architecture:

Voice message received
→ Download .ogg file from Telegram API
→ Transcribe (speech-to-text)
→ Feed transcript into existing anti-spam pipeline
→ Same AI context analysis as text messages
→ Decision: ban / warn / allow

Key decisions:

We transcribe everything under 5 minutes (covers 99% of spam voice notes)
Transcription runs async — doesn't block the moderation pipeline
The transcript gets the same 8-layer analysis as text: whitelist → global ban → reputation → trust → fingerprint → rules → AI context → decision
Language detection handles Russian, Ukrainian, and English voice messages
The result: a 15-second voice note saying "free betting tips, guaranteed profit" gets transcribed, classified as gambling spam, and the user gets banned — all within 3-5 seconds.

Image Spam Detection
Architecture:

Image received (photo or document)
→ Send to Vision AI
→ Analyze: is there promotional/spam text in the image?
→ If spam detected → classify category → ban
→ If clean → allow

What Vision AI catches:

Screenshots of fake profit/portfolio charts with channel links
Photos with overlaid text advertising gambling/crypto
Casino/betting ad graphics
Profile avatars that are literally advertisements
We only trigger Vision analysis on images from untrusted users (trust score below threshold). Trusted members' images pass through without Vision analysis — saves cost and reduces latency.

Anti-Masking
Spammers discovered they could bypass keyword filters with tricks like:

З🎰а🎰р🎰а🎰б🎰о🎰т🎰о🎰к (emoji between letters)
3аработок (number 3 instead of letter З)
Зaрaботок (Latin 'a' instead of Cyrillic 'а')
Our approach:

Raw message text
→ Strip emoji and special characters
→ Normalize Unicode (Cyrillic/Latin homoglyphs)
→ Normalize number→letter substitutions
→ Feed cleaned text into AI analysis
→ AI evaluates meaning, not characters

The AI layer is the key — even after normalization, context matters. "Заработок" in a freelance group is normal. "Заработок" in a cooking group is spam. Same word, different context, different decision.

Updated Pipeline (10 layers)

Whitelist check → free, 0ms
Global ban check → free, 0ms
Reputation auto-ban → free, 0ms
Trust system check → free, 0ms
Anti-masking normalization → free, 1ms
Fingerprint matching → free, 1ms
Rule-based detection → free, 0ms
Voice transcription → if voice message
Vision AI analysis → if image from untrusted user
AI context analysis → for edge cases
Decision → ban / mute / allow

Cheapest checks first. AI and Vision only for what cheaper layers can't decide.

Results
Voice spam: from 0% detection → ~95% detection
Image spam: from 0% detection → ~90% detection
Masked text: from ~60% → ~95% detection
Overall accuracy: maintained at 99.7%
False positive rate: still near zero
What's Next
Video message analysis is on the roadmap. Spammers will try video next — we'll be ready.

→ personym-ai.com/moderator-ai
→ Try free for 7 days

How We Increased Telegram Channel Comments by 340% in 2 Weeks (Real Examples)

PersonymAi — Thu, 26 Mar 2026 08:34:40 +0000

Every Telegram admin knows the feeling. You spend hours creating a quality post. You publish it. And then... silence. Zero comments. Maybe one emoji reaction.

Your subscribers see the silence too. And they leave.

This is the cold-start problem — and it kills more Telegram channels than bad content ever will.

The Death Spiral

Here's what actually happens:

New post → 0 comments

Subscribers see empty comment section → "dead channel"

They don't engage → less reason for others to engage

Growth stalls → subscribers slowly leave

We've seen channels with 10,000+ subscribers averaging 0-2 comments per post. The content was great. The engagement was dead.

What We Tested

We connected PersonymAI to 5 channels in different niches (crypto, news, tech, motivation, entertainment) and tracked results for 2 weeks.

Setup took 10 minutes per channel:

Connected the bot

Selected persona types that match the niche

Set comment frequency and timing

Done

The Results (2 weeks)

Metric Before After Change

Avg. comments per post 1.3 5.7 +340%

Unique real commenters per week 4 17 +325%

Avg. time to first comment 4+ hours 2 minutes -99%

Subscriber retention rate 82% 94% +12%

Why It Works

The key insight: real people comment more when they see others commenting.

Nobody wants to be the first commenter on a silent post. But when there's already a discussion happening — arguments, jokes, reactions — people naturally want to join in.

That's exactly what our AI personas do. They don't post "Great content! 👍". They have actual opinions:

Crypto channel example:

Persona "Skeptic": "BTC at 95K and people are still calling for 150K? Show me the on-chain data."

Persona "Degen": "95K is nothing, we're going to 200K easy 🚀🚀🚀"

Persona "Analyst": "Support at 92.5K looks solid. If we hold above 94K on the daily close, 100K is realistic."

Real subscriber joins: "Analyst is right, the RSI is showing..."

That last line is the magic. A real person saw a discussion and wanted to participate. That would never happen under a silent post.

What Makes Personas Feel Real

Each persona has a consistent personality that evolves over time

They argue with each other (not just agree)

35%+ of comments are short (1-5 words) like real chat

They use stickers, GIFs, and reactions

They reply in threads, not just flat comments

They reference the actual post content, not generic reactions

The Unexpected Effect

After about a week, something interesting happened across all 5 channels: organic comments started growing on their own. Even on posts where AI personas hadn't commented yet, real subscribers were more active than before.

Why? Because the channel no longer felt dead. Subscribers had been retrained to expect discussions. The culture changed from "lurking" to "participating."

Niche-Specific Results

Niche Best persona types Avg. organic comment growth

Crypto/Trading Analyst + Degen + Skeptic +420%

News Hot-take + Moderate + Cynic +280%

Tech Builder + Critic + Enthusiast +310%

Motivation Supporter + Realist + Joker +190%

Entertainment Memer + Troll + Fan +380%

Crypto performed best because the audience is already opinionated — they just needed a spark.

For Admins Who Also Deal With Spam

Half of these channels also had a spam problem. While testing the comment system, we also ran ModerAI anti-spam on the groups.

Results:

99.7% spam caught

0 false bans in 2 weeks

Admins spent zero time on moderation

ModerAI has a 7-day free trial — connect it and forget about spam.

How To Try It

AI Comments: Check plans at personym-ai.com — connect your channel and see first comments within minutes.

ModerAI Anti-Spam: personym-ai.com/moderator-ai — 7-day free trial, no credit card.

If you're an admin who cares about your community — you know the difference between a dead channel and a living one. We help you make that switch.

→ personym-ai.com

Metric	Before	After	Change
Avg. comments per post	1.3	5.7	+340%
Unique real commenters per week	4	17	+325%
Avg. time to first comment	4+ hours	2 minutes	-99%
Subscriber retention rate	82%	94%	+12%

Niche	Best persona types	Avg. organic comment growth
Crypto/Trading	Analyst + Degen + Skeptic	+420%
News	Hot-take + Moderate + Cynic	+280%
Tech	Builder + Critic + Enthusiast	+310%
Motivation	Supporter + Realist + Joker	+190%
Entertainment	Memer + Troll + Fan	+380%

From Stealth to Launch: How We Built AI Tools for Telegram (and Why We're Going Public Today)

PersonymAi — Thu, 26 Mar 2026 07:48:34 +0000

For over a year, our company didn't have a public website, a Twitter account, or a Product Hunt page. We were building AI tools for Telegram under NDA with international clients.

Today we're launching publicly. Here's what we learned.

The Technical Challenge
We built two products:

AI Comment System — generates natural Telegram discussions using 1,000+ persistent AI personas.

The hard part wasn't generating text. It was making it feel human. We solved this with:

Multi-pass quality pipeline (generate → self-check → enforce style)
Opinion Drift — personas change views gradually over time
65-85% threaded replies (real conversations, not flat comments)
35%+ short comments (1-5 words, like real chat)
Typing emulation and post-reading behavior

ModerAI Anti-Spam — context-aware spam detection for Telegram groups.

Traditional bots use keyword matching. We use an 8-layer pipeline:

Whitelist/ban check (free, instant)
Reputation scoring (3+ bans = auto-ban)
Trust system (skips 90-95% of legit messages)
Fuzzy text fingerprinting
39+ rule-based patterns
AI context analysis (only for edge cases)
Final decision
Result: 99.7% accuracy, near-zero false positives.

What We Learned
NDA work is great for building, terrible for marketing. We had a proven product and zero public presence. Starting from scratch on SEO, social, and community after a year of silence is harder than building the tech.

Keyword-based spam detection is fundamentally broken. The same word means different things in different contexts. AI context analysis isn't just better — it's a different category.

Persistence matters more than personality in AI personas. The biggest unlock wasn't making each persona unique — it was making them consistent over time.

Launch Day
We're live on Product Hunt today: https://www.producthunt.com/posts/personymai/maker-invite?code=wMwcDO

If you're interested in Telegram tooling, AI-powered content generation, or anti-spam systems, I'm happy to go deeper on any of these topics.

→ personym-ai.com

Lessons From Processing Millions of Telegram Messages: What We Learned About Spam

PersonymAi — Wed, 25 Mar 2026 10:18:09 +0000

We've spent years building an AI anti-spam system for Telegram. After processing millions of messages across hundreds of communities, here are the patterns we discovered.

Spam Has Evolved

Forget the obvious stuff — links, ALL CAPS, "CLICK HERE FOR FREE MONEY."

Modern Telegram spammers are sophisticated:

The Edit Trick
Post a normal message. Wait an hour. Edit it into a scam link. Most bots never re-check edited messages. We learned this the hard way and built edit monitoring into our core pipeline.

The Trust Builder
Join a group. Post 5-10 normal messages over a few days. Build credibility. Then drop the spam. Keyword filters can't catch this because the spam message itself might look innocent — it's the pattern that's suspicious.

The Avatar Bait
Create accounts with provocative profile photos. Join groups. Post nothing — the avatar itself is the spam (drives clicks to the profile with links in bio). This requires pre-message analysis that most bots don't do.

The Multi-Account Wave
Hit a group with 20 different accounts in 5 minutes. Even if the admin bans them, the damage is done — members saw the spam. Speed of response matters more than accuracy here.

What Keyword Filtering Gets Wrong

We analyzed false positive rates across traditional moderation bots. The results were painful:

The word "investment" triggers bans in 73% of keyword-based bots. But in crypto and trading groups, it's used in normal conversation hundreds of times per day.

"Free" is flagged by 61% of bots. But "free trial", "free tier", and "free update" are perfectly legitimate.

The fundamental problem: context determines whether a message is spam, not individual words.

The Global Network Effect

Our biggest insight came from connecting multiple chats into a shared ban network.

When Chat A bans a spammer, Chats B through Z know about it instantly. The spammer can't just move to the next group.

After connecting 100+ chats:

New spam accounts were blocked on first appearance in 89% of cases
The average time to neutralize a spam wave dropped from 15 minutes to under 30 seconds
False positive rate decreased as more data flowed through the network

The network gets smarter with every chat added. It's not linear growth — it's exponential.

Trust Is Better Than Rules

Early versions of our system were too aggressive. We caught spam, but we also annoyed legitimate users.

The breakthrough was shifting from "block suspicious behavior" to "build and track trust."

Every user in our system has a trust score based on:

Message count and quality
Behavior consistency over time
Reputation across the network
Account age and profile completeness

High-trust users are never bothered. New users get gradually more freedom. Spammers never build enough trust to bypass the system.

This reduced false positives to near zero while maintaining 99.7% detection accuracy.

Fingerprinting Beyond Accounts

Banning an account is easy. Banning a person is hard.

Spammers create new accounts constantly. But their behavior patterns are remarkably consistent:

Message timing intervals
Text structure and formatting habits
Target selection patterns
Time-of-day activity profiles

Our fingerprint system identifies these patterns even across completely new accounts. It doesn't matter if the username and phone number are different — the behavior signature matches.

Numbers After Years of Production

99.7% spam detection accuracy
~0% false positive rate
Sub-second average decision time
Millions of messages analyzed
Hundreds of active communities protected

What's Next

Spam evolves constantly. We're working on:

Voice message spam detection
Image and media content analysis
Predictive blocking (identifying potential spammers before they act)
Cross-platform intelligence sharing

The arms race never ends, but with AI that understands context and a network that shares intelligence, the defenders finally have the advantage.

ModerAI is part of PersonymAI. If you manage Telegram communities and want to test it: personym-ai.com — 7 days free.

Questions about our architecture or approach? Happy to discuss below.

Building AI Personas That Sound Human: Our Approach to Telegram Engagement

PersonymAi — Wed, 25 Mar 2026 10:15:34 +0000

Dead comment sections kill Telegram channels. You post great content — zero reactions. New subscribers see silence and leave.

We solved this by building AI personas that engage like real people. Here's how.

The Problem With Generic AI Comments

Everyone has seen them: "Great post! Thanks for sharing! Very informative!"

These comments are worse than no comments at all. They scream "bot" and destroy trust.

Real Telegram chats look nothing like this. Real people write "lol", argue with each other, use slang, drop stickers, and type one-word reactions.

Our Persona Architecture

Every AI account in PersonymAI has a persistent identity:

Personality traits:

Writing style — formal, casual, slang-heavy, emoji addict, minimalist
Opinion pattern — bullish, bearish, contrarian, neutral
Aggression — scale of 0 to 100 (polite analyst to aggressive degen)
Language — strict Ukrainian, Russian, surzhyk, or mixed

Behavioral rules:

Each account has a unique typing speed
Some accounts comment early, others are late reactors
Some reply more than they initiate
Sticker usage varies (15% of comments include stickers)

Two accounts never produce the same output for the same input. Ever.

Opinion Drift

Static personas feel fake within a week. Real people change their minds.

We implemented Opinion Drift — accounts gradually shift their positions over time. A bullish account won't suddenly turn bearish overnight. Instead, sentiment shifts slowly based on market conditions and community reactions.

This creates realistic long-term behavior that's indistinguishable from organic users.

The 3-Pass Quality Pipeline

Raw AI output is never good enough. Our pipeline:

Pass 1: Generation
Our proprietary AI generates comments using the post content, channel niche, and persona profile. A 1400+ line prompt system handles persona injection and niche-specific terminology.

Pass 2: Self-Check
Two groups of 5 validation rules:

Group A: topic relevance, persona consistency, language correctness
Group B: length check, style verification, repetition detection

Pass 3: Short Enforce
Real chats have lots of ultra-short messages. We enforce that 35%+ of comments are 1-5 words. Making AI write "lol" instead of a paragraph is surprisingly hard.

The Cleanup Layer

Even after 3 passes, bot patterns leak through:

Unnecessary dashes and formal punctuation → removed
English slang in Russian comments → caught and fixed
ALL CAPS overuse → toned down
"I've been in crypto since 2021" → classic AI pattern, blocked
Analytical tone from a "degen" persona → detected and rewritten

Threading: The Secret Sauce

65-85% of our comments are threaded replies. Accounts don't just comment — they argue with each other, agree, joke, and create sub-discussions.

A flat wall of independent comments looks artificial. A thread where someone says "you're wrong" and gets three different responses looks real.

Real-Time Context

For crypto channels, our system integrates live market data — prices, volumes, 24h changes. Comments reference actual numbers and current events.

"BTC just broke 67k" hits different than "the market is showing positive movement."

Results

Channels using our system report:

300% increase in organic engagement within the first month
40% higher subscriber retention
Natural-looking discussions from day one

The bridge between an empty channel and a thriving community is shorter than you think.

We're always looking for feedback from the dev community. What would you want to see in an AI engagement system?

PersonymAI