Skip to main content
This page defines the core abstractions that make up Lola Send’s architecture. Understanding these primitives is essential for evaluating how the platform integrates with a financial institution’s existing infrastructure and compliance requirements.

​
Core abstractions

MessageGateway — The central HTTP server that receives messages from channel connectors, runs them through the middleware pipeline, and dispatches to the appropriate agent. MessageGateway is the single entry point for all inbound messages, regardless of channel. It enforces middleware ordering and ensures every message passes through authentication and enrichment before reaching an agent. Celai — Leap Financial’s proprietary agentic framework. Celai provides the foundational components that Lola Send is built on: MessageGateway, channel connectors, LogicRouter, MacawAssistant, the middleware chain, and Redis-backed state and history stores. It is a general-purpose framework for building conversational AI applications on messaging channels. MacawAssistant — The agent implementation that combines a system prompt, AI model configuration, and a set of tool functions into a self-contained conversational unit. Each agent in Lola Send is a MacawAssistant. The bank controls agent behavior by modifying the system prompt and the set of available tool functions — no application code changes are required. LogicRouter — The component that evaluates sender state and compliance status to select the right agent for each conversation. LogicRouter directs messages to one of five agents:
  • Onboarding agent — unidentified senders; guides first-time verification
  • Sender home agent — verified senders; full conversational capabilities
  • Blocked agent — senders flagged by compliance; restricted interaction
  • Pending CIP agent — senders with in-progress verification; limited interaction
  • Service unavailable agent — graceful degradation when backend services are down
Tool function — A backend service invocation declared on an agent. Agents call tool functions for all factual data — quotes, recipients, operations, banks, countries. The AI model never generates this data; it only decides which tool function to call and how to present the result. This architecture eliminates hallucination for financial data and ensures every data point traces back to a backend service. Middleware — An ordered pipeline of processors that authenticate, enrich, and transform messages before they reach an agent. Each middleware stage runs in sequence: authentication, contact normalization, command handling, and Chatwoot synchronization. The bank controls which middleware runs and in what order. Connector — A pluggable adapter for a messaging channel. Connectors normalize inbound messages from a specific platform (WhatsApp, Telegram) into Celai’s internal format and format outbound responses for that platform’s API. Adding a new channel requires implementing a connector — agent logic remains unchanged. Conversation state — Redis-backed session data that persists across messages within a conversation. The application itself is stateless; all session context lives in Redis, and identity data lives in backend services. This design enables horizontal scaling and ensures no sensitive data persists in the application layer. Escalation — The process of routing a conversation to a live human agent via Chatwoot when it falls outside automated handling. The bank defines escalation triggers. Escalation preserves the full conversation history, giving the human agent complete context.

​
Terminology reference