GroupShield
Managing a WhatsApp group shouldn't be a constant battle.
GroupShield enforces your rules for you - automatically.
GroupShield
Managing a WhatsApp group shouldn't be a constant battle.
GroupShield enforces your rules for you - automatically.
One bot to rule them all. Simple for admins, transparent for users.
Smart two-layer detection catches offensive words even when users try to bypass filters with letter substitutions or creative misspellings. Zero false positives — the bot only acts when it's certain.
Warn, Delete, Remove, Block. You define the escalation path that fits your community culture.
No complex dashboards. Configure rules, languages, and enforcement steps through a natural DM conversation.
Block forbidden phrases or enforce "Allowed Only" lists. Supports exact matches and regex containment.
Define precise time windows when messages are allowed in your group. Shabbat & holiday mode is also built in — the group locks automatically before Shabbat and every Jewish holiday, and reopens when they're over.
Full Hebrew & English support, a dedicated management group for your admins, automatic group description sync, and a smart welcome message for new members.
Real data from the running production instance.
A visual simulation of how GroupShield transforms your community management.
Incoming message is captured directly from the managed group.
Cross-references content against regex logic and configured spam limits.
Executes strict enforcement: deletes violation, warns, blocks or removes.
Delivers transparent logs to the admin group and a private DM to the user.
A real-time trace of a message engineered to slip through every filter.
The forbidden-word list is checked, including common variations and alternate spellings. "ז0נה" doesn't appear in the list — the digit '0' instead of 'ו' breaks the match.
The message is converted to a bigram vector and compared against all words in the list. Best score: 0.09 — well below the 0.65 suspicious threshold. No match.
GroupShield scans for evasion patterns. Pattern 4 fires: digit substitution detected — '0' replacing 'ו' in a known offensive root. The message is flagged and routed to AI for review.
The model receives the message with a zero-tolerance system prompt. It returns YES in under 4 seconds — Arabic-origin insult confirmed. Result is cached to prevent duplicate API calls.
The violation triggers the enforcement pipeline: message deleted from the group, private DM warning sent to the user, then removal. Each step is logged with a unique action ID for potential undo.
The full enforcement report is sent to the management group. The LLM catch is flagged — the developer reviews the pattern and decides whether to add "ז0נה" to the permanent word list so future instances are caught instantly, without needing the AI.
For the technically curious — the architecture powering GroupShield.
Reliable WhatsApp Web automation utilizing persistent sessions and a self-healing crash recovery protocol.
Local persistent database structured to handle configurations and warnings for thousands of groups efficiently.
Rate-limited message processing pipeline that prevents connection bans and ensures deterministic rule evaluation.
Containerized setup for effortless deployment across any cloud provider with zero dependency conflicts.
Advanced crash-recovery system that automatically restores the WhatsApp connection if the browser unexpectedly drops.
Configure complex rules, schedules, and immunity groups intuitively inside WhatsApp. No dashboard required.
The AI moderation layer runs Llama-3.1-8b-instant via Groq with a 4-second timeout and a monthly usage cap. A suspicion scorer pre-filters messages before any API call. Prompt injection attempts — in Hebrew or English — are blocked instantly without involving the LLM at all. Fails silently on infra errors: zero false positives.
Every Thursday a cron job fetches candle-lighting times (Jerusalem) and Havdalah times (Netanya) from the HebCal API. The bot locks groups 5 minutes before entry and unlocks 5 minutes after exit, re-locking on any manual override. If the fetch fails, a guided recovery flow lets the admin enter times manually.
graph TD
A["📱 WhatsApp Group"] -->|"New Message"| B["🤖 Puppeteer Bot\n(Node.js)"]
B -->|"Enqueue"| C["⚡ Event Queue\n(Rate-Limiter)"]
C --> D["🧠 Rule Engine\n(Regex / Time / Spam)"]
D -->|"Clear Violation"| E["⚖️ Enforcement Pipeline"]
D -->|"Suspicious / Context Word"| LLM["🤖 AI Layer\n(Groq / Llama 3.1)"]
LLM -->|"Violation"| E
LLM -->|"Clean"| F["✅ Ignore"]
D -->|"Immune user"| F
E --> G["🗑️ Delete Message"]
E --> H["📩 DM Warning"]
E --> I["🚫 Remove User"]
E --> J["📣 Report to Mgmt Group"]
SCHED["🕯️ Shabbat Scheduler\n(Cron / HebCal API)"] -->|"Friday evening"| LOCK["🔒 Lock Group\n(Admins Only)"]
SCHED -->|"Saturday night"| UNLOCK["🔓 Unlock Group"]
B -->|"Read/Write"| K[("🗄️ SQLite DB\n(Multitenant)")]
K --- L["Groups Config"]
K --- M["Warnings Log"]
K --- N["Immune List"]
style A fill:#075e54,color:#fff
style B fill:#128c7e,color:#fff
style C fill:#1a1a2e,stroke:#00e5ff,color:#fff
style D fill:#1a1a2e,stroke:#00e5ff,color:#fff
style E fill:#1a1a2e,stroke:#f59e0b,color:#fff
style LLM fill:#1a1a2e,stroke:#a855f7,color:#fff
style SCHED fill:#1a1a2e,stroke:#f59e0b,color:#fff
style LOCK fill:#1a1a2e,stroke:#ef4444,color:#fff
style UNLOCK fill:#1a1a2e,stroke:#00ffa3,color:#fff
style K fill:#1a1a2e,stroke:#00ffa3,color:#fff