01
Engagement Channels
Web console, WhatsApp messaging and real-time voice all feed the same conversation pipeline.
Operator web consoleWhatsApp gatewayReal-time voice layer
System Architecture
A grounded AI pipeline behind every channel.
Distributors reach the platform through web chat, WhatsApp or voice. Each conversation flows into one backend, through one AI reasoning core, and onto a shared data layer — so behaviour, data and audit trail stay identical everywhere.
Business Control Plane
Ops & campaign management
Operator Console
Audience, schedule, conversation review
Call Orchestration
Outbound queue, retries, suppression
Analytics Dashboards
Outstanding, ageing, collection outcomes
Team & Access
Users, roles, depot assignment
Real-Time Voice Layer
Telephony & speech I/O
Telephony Gateway
Cloud telephony · SIP · PSTN
STT + Language Detection
Noise, interruption, IVR handling
TTS Voice Response
Natural multilingual speech
Agentic AI Brain
Orchestration & reasoning
LLM Orchestrator
Hosted reasoning model
Conversation Policy Engine
FAQ & collection flow, guardrails
Structured Extraction
Intent, objection, dispute fields
Tonality Controller
Warm, official, persuasive
Conversation Memory
Current call + prior interactions
Tool Router
Approved business actions only
Dealer + Receivables Intelligence
Context & knowledge data
Dealer Profile Store
Category, market, depot, territory
Outstanding & Ageing Engine
Balances, buckets, due dates
Debit-Note Engine
DN stages, rates, early-pay savings
RAG Knowledge Layer
Schemes, FAQs, objection handling
Collection Outcome Workflow
Capture → classify → hand off
Intent Classification
YES · NO · DEFER · DISPUTE
Outcome Logging
Structured record per conversation
Dispute & Escalation
Depot hand-off and support request
Enterprise Integrations
Messaging & backend systems
Messaging Gateway
WhatsApp two-way + templates
Email & Reminders
Automated debit-note reminders
Data Ingestion & Warehouse
Receivables sync + analytics store
Observability & Audit
Trace every message & AI turn
Immutable Message History
Every message, every channel, timestamped
Raw AI Audit Log
Full prompt + response stored per turn
Conversation Replay
Turn-by-turn reconstruction on demand
Outcome Ledger
Structured intent recorded per conversation
Error & Trace Monitoring
Live capture, alerting, root-cause
Layers from business control at the top down to enterprise systems — every channel flows through the same AI brain and shared data.
Layers
Four cooperating layers.
02
Application Backend
A single API service handles authentication, routing, real-time sockets and orchestration for every channel.
REST + WebSocket APISession & access controlChannel webhooks
03
AI Agent Core
The reasoning engine selects data tools, retrieves grounded facts, and composes the reply — every answer is backed by real data.
Tool-required reasoningFAQ & Collection modesDeterministic classification
04
Data & Analytics
Operational records, fast analytics, low-latency cache and secure file storage underpin the platform.
Operational databaseAnalytics warehouseSession cacheSecure file store
Request lifecycle
From inbound message to logged outcome.
- 01
Capture
A message arrives from any channel and is persisted to history and cache.
- 02
Ground
The AI core is required to call data tools — it cannot answer without real records.
- 03
Reason
Tool outputs are fed back for a second pass that composes a concise, channel-appropriate reply.
- 04
Act
Collection outcomes are logged and disputes are escalated to the depot team automatically.
Design principles
The decisions that keep it trustworthy.
Grounded by design
The reasoning layer must call a data tool on every turn. Outstanding amounts and invoice details can never be fabricated.
Two-pass prompting
Pass one retrieves data; pass two presents it. Retrieval logic stays clean and presentation can be tuned per channel.
Deterministic outcomes
Classification runs at zero temperature, so the same dealer statement always yields the same structured result.
Single source of truth
Every channel reads and writes through the same backend, so web, messaging and voice share one audit trail.
Access at the prompt
Dealer-versus-admin data boundaries are enforced in the agent's instructions as well as the API authorization layer.
Auditable by default
Each interaction stores its raw request, response and final reply for full replay and compliance.
Auditability
Every message and AI turn is fully traceable.
Nothing the agent says or decides is a black box. Each interaction writes an immutable record — the inbound message, the exact prompt sent to the model, the model's raw response, the reply delivered, and the structured outcome. Any conversation can be replayed end to end for dispute resolution, compliance and quality review.
- Immutable history — records are appended, never overwritten
- Raw AI request & response captured for every turn
- Outcome ledger ties each conversation to a structured intent
- Scoped by dealer, agent, channel and timestamp
Captured for every interaction
- 1
Inbound
Message captured from any channel with sender, mode and timestamp.
- 2
Persisted
Written to immutable history and recent-conversation cache.
- 3
AI request
The exact prompt and context sent to the model are stored.
- 4
AI response
The model's raw response and any tool calls are stored.
- 5
Reply logged
The final reply delivered to the dealer is recorded verbatim.
- 6
Outcome
Collection intent and escalations are written to the ledger.
Data model
Core entities at the heart of the platform.
payment_followup Invoice & transaction lines — the core receivables data. outstanding_summary Aggregated balances and ageing buckets per account. customer_master Credit terms, debit-note stages and account status. collection_intent Structured outcome of every collection conversation. message_history Full, immutable record of all channel messages. message_audit Raw AI request, response and final reply per turn — the audit trail. user_concern Disputes and escalations routed to the depot team.