Target architecture
One shared context, multiple model paths.
1 · Input
Customer messagepolicybazaar.aiWhatsAppPB BuddyWebsite
2 · State
Conversation backend
Loads the customer and what they last asked, before every turn
Customer facts
Per person, every product
Journey facts
Per journey, resumable
3 · Routing
Backend rules
Confirmations; no model call
Router · small model
Everything else; under 50 ms added
4 · Models
PB Flash
Details, quotes, forms; most turns
PB Pro
Plan fit, clauses; hard turns only
Product skills
HealthTermMotor+4 more
5 · Output
Answer to the customer + proposed memory updates
The backend checks whose fact it is before saving
Audit trail
Every turn, with its trace id
6 · Quality
Evals on every answer
- Premiums traced
- No claim promise
- IRDAI disclosures
- Same language
- No personal data
- Injection logged
Observability dashboard
Quality
98.5%
Cost/turn
0.4 p
p95
1.2 s
Blocked
12
Overview
One shared context, multiple model paths
Every customer turn takes the same path: it is read in, given the customer's context, routed, answered by the right model, and checked. Click to follow one message through.
Click the diagram, or use the arrow keys