BrokerBee AI Platform

BrokerBee is an AI-powered real estate concierge platform built on GoHighLevel (GHL). Its core product, "Barry", is an AI agent that handles inbound Voice and SMS/Chat conversations to intelligently route callers to the correct real estate vendor or partner agent based on their service category and location. The platform serves real estate brokers and their vendor networks, enabling fast, accurate, and branded customer interactions without manual intervention.

The existing GHL environment suffered from several critical issues: a routing bug causing misdirected calls (e.g., Alabama routing failures for agent Brooke Brunsvold), inconsistent Voice vs. Chat prompts with hardcoded agent names, and a probabilistic Knowledge Base (KB) that produced unreliable vendor lookups. The absence of an automated testing framework meant regressions went undetected, and there was no structured process for non-technical team members to onboard new agents or vendors.

Platform: GoHighLevel (GHL)

Retrieval Service: Node.js REST API + PostgreSQL

Testing Framework: Promptfoo

Authentication: Bearer Token

Project Management: Atlassian Jira

A two-phase delivery was executed. Phase 1 stabilized the existing GHL system, auditing the environment, canonicalizing the AI prompt, and deduplicating the Knowledge Base. Phase 2 replaced the GHL KB with an external deterministic PostgreSQL retrieval service, providing reliable, schema-enforced vendor lookups registered as a GHL Tool.

Phase 1: Stabilization A forensic audit of the GHL environment was performed, exporting all Knowledge Base entries and mapping existing agent-to-vendor relationships. A single "Canonical Prompt" was developed to replace divergent Voice and Chat prompts, using GHL Custom Values for channel-toggling and eliminating all hardcoded agent names. The Knowledge Base was deduplicated with vendor phone numbers verified in E.164 format. A Promptfoo test suite of 20-30 golden-path scenarios was implemented, covering Agent Identification, Vendor Retrieval, and Out-of-Scope Handling, targeting a 100% pass rate on critical scenarios.

Phase 2: External Retrieval Service A live REST API endpoint (POST /lookup-vendor) was deployed backed by a PostgreSQL database, exposing a strict JSON schema (Name, Phone, Category, AgentID) secured with Bearer Token authentication. All 100+ vendor records were migrated via an ETL pipeline with zero data loss, normalizing service categories (e.g., mapping "AC Repair" and "HVAC" to a single category). The retrieval service was registered as a Tool in GHL, with the Canonical Prompt updated to invoke it instead of native KB search. An expanded evaluation harness of 100+ scenarios, including adversarial tests, validated 100% deterministic results. A phased rollout across all partner agents achieved a <1% error rate over 48 hours. A Barry Admin Runbook was delivered, enabling new agent onboarding in under 5 minutes.