PRODUCTION MULTI-AGENT ARCHITECTURE • FULL-STACK AI SYSTEM

CyberSafe AI — Agentic AI-Powered Cyberbullying Detection Platform

An enterprise-grade, privacy-preserving AI software system orchestrating an autonomous 7-agent pipeline, fine-tuned RoBERTa transformer inference, dense RAG knowledge retrieval, real-time affective risk matrices, and global conversational safety support.

FastAPI (Python 3.9+)RoBERTa Transformer ActiveNext.js 16 (App Router)7-Agent OrchestratorMiniLM Dense Vector RAGDifferential Privacy & PII RedactionZero-Downtime Fallback
CENTRAL APPLICATION HUB

Test the Multi-Agent Detection Pipeline in Real Time

Evaluate realistic sample texts, inspect the full 7-agent trace, observe affective distribution, and test the persistent AI assistant copilot.

Launch DetectorAnalytics DashboardAbout the System
Active Live Model
RoBERTa Detection
Deep bidirectional sequence classification analyzing 512-token context windows for subtle cyberbullying nuances and harassment.
Launch Detector
7 Autonomous Agents
Multi-Agent DAG
Stateful orchestration coordinating PII sanitization, RoBERTa inference, affect NLP, context nuance, and deterministic risk scoring.
Open Dashboard →
Zero-Leakage Privacy
Privacy & Grounding
Automatic regex redaction for all personally identifiable info paired with dense vector RAG knowledge retrieval and prompt injection shields.
About the System →
cybersafe-pipeline-orchestrator.ts • live inference telemetry
HEALTHY • 0.0% ERROR
TOTAL PIPELINE: 78ms
Input IngestionSanitized
"You are worthless and nobody likes you."
PII Check: Clean (0 hits)Tokens: 8
RoBERTa Primary Model
Cyberbullying
Confidence:75.6%
Status:Active
Dominant Affect:Anger (31.2%)
Agentic Risk & Grounding
MEDIUM RISK

XAI Explanation: Hostile lexical patterns detected. RAG cited platform evidence preservation protocol.

7 Agents Executed Sequentially
AGENTIC EXECUTION FLOW

Concise Visual Architecture: The 7-Agent Autonomous Pipeline

Every message passes through a deterministic, high-throughput multi-agent DAG that decouples detection, affect, contextual threat, deterministic risk gating, and verified RAG knowledge.

Interactive Pipeline Diagram (Click any step to inspect responsibilities)Total SLA: < 100ms
▶
▶
▶
▶
▶
▶
Stage 01: Detection Agent
RoBERTa Sequence Classifier
Active Live

Executes deep bidirectional transformer classification on the 512-token input sequence for hostility and cyberbullying.

Expected Latency:
~ 35ms
Detection Agent→Emotion Agent→Context Agent→Risk Agent→RAG Agent→Explanation Agent→Response Agent
PRODUCTION-ORIENTED CAPABILITIES

Real AI Software Engineering & Systems Integration

Engineered beyond simple wrapper scripts. Built with real asynchronous micro-services, transformer sequence modeling, vector retrieval, and zero-downtime edge redundancy.

Active Live Model
RoBERTa — Active Primary Model
Fine-tuned deep bidirectional transformer engineered for high-precision sequence classification of cyberbullying and online hostility.
Sequence length: 512 tokens
CardiffNLP twitter-roberta base
Calibrated probability outputs
Sub-100ms multi-agent DAG latency
7 Autonomous Agents
Autonomous Multi-Agent Architecture
Stateful agentic orchestration coordinating detection, affect, context, risk, RAG retrieval, and explanation in an asynchronous DAG.
Asynchronous task chaining
Typed CyberbullyingState schema
Individual agent latency profiling
Fault-tolerant fallback paths
5 Emotional Classes
Affective Emotion & Sentiment Vector
Granular emotional profiling distinguishing anger, fear, sadness, neutral, and joy alongside VADER compound sentiment.
Multi-class probability distribution
Hostile emotional surge detection
Sarcasm & tone divergence flags
Psychological distress indicators
MiniLM Cosine Similarity
Dense Vector RAG Knowledge Base
Grounding moderation and chatbot responses in verified literature on digital civility, evidence preservation, and crisis hotlines.
all-MiniLM-L6-v2 embeddings
Exact dot-product cosine ranking
Chunked markdown knowledge docs
Citations in every API response
PII Redacted + Anti-Jailbreak
Enterprise Privacy & Security
Zero raw PII storage, regex identity anonymization, input sanitization, and adversarial prompt injection interception.
Email, phone, IP & URL redaction
Heuristic prompt injection shields
SQLite local encrypted persistence
Zero data leak to external third parties
FastAPI + Next.js Edge
Production Full-Stack Engineering
Decoupled architecture pairing a high-throughput Python FastAPI backend with a Next.js App Router frontend and serverless fallbacks.
Pydantic v2 strict typing
Zero-downtime Vercel fallback
Sub-100ms pipeline execution
Standardized REST endpoints
FULL-STACK ARCHITECTURE

Decoupled FastAPI Engine + Next.js Edge

Production deployment architecture supporting both local Python execution and standalone Vercel deployments via resilient zero-downtime serverless analyzer fallbacks.

Backend Core & Machine Learning Engine
FastAPI + PyTorch + Transformers
High-throughput asynchronous service running on Python 3.9+ with persistent vector embeddings and conversation memory.
API Routes Specification:
POST /api/analyze • 7-Agent DAG Pipeline
POST /api/detect • Raw RoBERTa sequence inference
POST /api/chat • LLM + RAG Grounded Copilot
POST /api/rag/search • Dense MiniLM vector similarity
GET /api/health • System uptime & model status
SQLite Conversation MemoryEnabled
Vector Index FormatDense Cosine MiniLM JSON
Frontend Platform & Edge Fallback Engine
Next.js 16 + Turbopack + Tailwind
Modern App Router with client-side interactive telemetry, global floating assistant, and dual-mode resilient proxying.
Edge Redundancy Strategy:

Proxy routes attempt the external FastAPI server with a 3.5s timeout. If unavailable (e.g. running standalone on Vercel), execution gracefully shifts to built-in TypeScript serverless analyzer engines with zero downtime.

Global Assistant IconMounted on All Pages
Active Live ModelRoBERTa Active Model

Ready to Explore the Platform?

Test live text samples, inspect the 7-agent DAG execution trace, or explore our complete system architecture.

Launch DetectorDashboardAbout the SystemContact →
RoBERTa (Active Live Model) is the sole active detection model powering live inference across the 7-agent pipeline.