← Back to architecture

Prael AI Agent — Conversation & Orchestration Flow

Full agentic pipeline from user input through intent classification, sub-agent routing, LLM orchestration, guardrails, and response delivery

Input channels
User touchpoints
Mobile app (iOS/Android)
Web dashboard
WhatsApp / SMS
Voice (Twilio IVR)
↓
Input normalization
Speech-to-text (Whisper)
Language detection
Message dedup
UTF-8 sanitization
↓
Session management
Session ID generation
Auth token validation
Rate limiter (60 req/min)
Context window loader
Avg session: 4.2 turns · 18s latency
User state
KYC tier
Verified
Credit score
720
Active products
3
Transaction history (90d)
Preference profile
Intent classification & routing engine
NLU pipeline
Tokenization
Entity extraction (NER)
Sentiment analysis
Intent confidence scoring
Accuracy: 94.7%
Intent categories
send_money (38%)
check_credit (22%)
card_inquiry (15%)
account_mgmt (12%)
financial_advice (8%)
escalate_human (5%)
Routing decision
Confidence threshold (>0.85)
Multi-intent resolver
Disambiguation prompt
Fallback → general agent
Priority queue (VIP tier)
Routing latency: 45ms p95
LLM orchestration layer
Prompt assembly
System prompt (persona)
Context injection (RAG)
User history (last 10)
Tool definitions
Constraint rules
Model selection
Claude 3.5 (reasoning)
GPT-4o (generation)
Kredete LLM (credit)
Cost-tier routing
Latency budget check
Tool calling
get_balance()
initiate_transfer()
pull_credit_report()
get_fx_quote()
schedule_payment()
Execution control
Max retries: 3
Timeout: 30s
Token budget: 4096
Streaming enabled
Circuit breaker
Specialized sub-agents
Remittance agent
Corridor lookup
FX rate engine
Rail selection
Compliance check
ETA calculation
$500M+ processed
Credit agent
Score retrieval
Factor analysis
Action plan gen
Bureau reporting
Progress tracking
+58pt avg lift
Card agent
Eligibility check
Application flow
Limit management
Dispute handling
Rewards lookup
12K cards issued
Finance advisor
Spend analysis
Budget planner
Savings goals
Investment recs
Tax optimization
NPS: 72
Escalation agent
Sentiment trigger
Complexity score
Agent handoff
Context transfer
SLA enforcement
Escalation: 5.2%
Response generation & delivery
Output guardrails
Hallucination check
PII redaction
Toxicity filter
Compliance scan
Brand voice check
Response formatting
Markdown → rich text
Card/button rendering
Amount formatting
Multi-language (i18n)
Accessibility (WCAG)
Delivery
WebSocket push
Push notification
SMS fallback
Email digest
Webhook (B2B)
Confirmation flows
Transaction preview
Biometric auth (Face ID)
OTP verification
Double-opt confirm
Receipt generation
Feedback & continuous learning
User signals
Thumbs up/down
Conversation rating
Task completion
Retry detection
Quality metrics
CSAT score (4.6/5)
Resolution rate (89%)
Avg turns to resolve
Hallucination rate (0.3%)
Training pipeline
RLHF data collection
Prompt refinement
Fine-tune trigger
A/B test routing
Observability
Trace ID logging
Latency histogram
Error rate alerting
Cost per conversation
Context & memory
RAG retrieval
Query embedding
Vector similarity (cosine)
Top-K retrieval (k=5)
Re-ranking (cross-encoder)
Context window packing
Retrieval: 12ms p50
Knowledge bases
Product catalog
Compliance rules
FAQ / Help center
Rate sheets (live)
Corridor matrix
Conversation memory
Short-term (session)
Long-term (Redis)
Episodic (user history)
Semantic (embeddings)
External APIs
Credit bureaus
Payment processors
FX providers
KYC/AML services
Card networks
Safety layer
Jailbreak detection
Active
Prompt injection guard
Active
Topic boundary enforcement
Regulatory constraint engine