AgentLah

Channel-agnostic conversational agent framework

A channel-agnostic conversational agent framework, deployed and live — one codebase powers a WhatsApp property-sales assistant and the real-estate sales agent delivered for a freelance client. Property-listing RAG, knowledge-base RAG, human-takeover dashboard, and pluggable LLM providers.

Years
2026
Timeline
Mar 2026 — present (client engagement ongoing)
Role
Author · freelance developer
Status
Live
Stack
TypeScript, Node.js, Express, WhatsApp Cloud API, PostgreSQL, Anthropic, OpenAI, Google Gemini
Links

Problem

Real-estate sales conversations happen on WhatsApp, on web chat, over email — and a bot built for one channel is a rewrite for the next. Meanwhile the conversations themselves need real substance: listing knowledge, lead qualification, and a way for a human to take over the moment the AI is out of its depth.

Approach

AgentLah separates the agent from the channel: one core handles conversation memory, RAG, and tool use; thin adapters connect WhatsApp Cloud API, web chat, or direct API. Two retrieval layers — property-listing RAG and knowledge-base RAG — ground answers in the client's actual inventory.

LLM providers are pluggable (Anthropic / OpenAI / Google), so the same agent can be tuned for cost or quality per deployment. A human-takeover dashboard lets an operator watch conversations and step in live, and MCP tool-use connects the agent to CRM and scheduling.

Outcomes

  • Deployed and live at agent-lah.vercel.app.
  • v1 of an autonomous sales agent delivered to a real-estate client (under NDA); the engagement is ongoing.
  • The framework is open source and generalises beyond property — the channel-adapter pattern is the reusable part.
Wong Kai Shen, 2026. Built in Kuala Lumpur. Say hello.