WhatsApp Multi-Agent Ops

API Agent Demo

Overview

Constantly switching between SQL editors and internal dashboards just to answer simple ops questions got old fast. So I built a bot you talk to like a coworker: ask a question in plain WhatsApp chat, and it routes the request to the right specialized agent to actually execute it. Built early, when agentic AI tooling was still emerging, mostly as a hands-on way to learn the mechanics. Worth noting this predates standardized tool-calling protocols like MCP: the registries, routing, and context retrieval here are all hand-rolled, work a protocol and a platform like OpenClaw would now handle for me.

Highlights

• Multi-agent architecture using LangChain, Gemini, and DeepSeek for dynamic request routing and task delegation.
• Retrieval-Augmented Generation (RAG) with MiniLM embeddings to pull only relevant API/DB registries into context.
• Secure SQL execution restricted to predefined, human-reviewed queries with a 60-second confirmation window before anything runs.
• Self-healing validation-and-retry loops that catch and correct malformed structured outputs before execution.
• Runs entirely on free-tier models for zero-cost daily operation, with a simple swap-in path to paid models when needed.

Built with