Full Stack Developer & AI Engineer
True Mortgages
AI-powered mortgage admin portal — an AI brain trained on Xavier's voice that automates lead engagement across 19,700+ contacts
Overview
True Mortgages is a centralized AI admin portal built for XA Financial Services — a US mortgage brokerage. The system ingests all existing GHL call recordings, SMS threads, and email conversations into a RAG pipeline, trains on the owner's communication style, and autonomously handles lead engagement via SMS and email. The portal gives the team a single command center: a real-time dashboard with 19,700+ active leads, an AI chat interface for natural language business queries, an activity feed showing every AI-generated message with confidence scoring, and a full data ingestion pipeline that sequences contacts → conversations → messages → transcriptions → vectorization → knowledge base.
The Challenge
Xavier Alcala's mortgage business had accumulated 15,000+ existing leads and a 100,000-prospect outreach list — all managed across GoHighLevel sub-accounts with no automation layer. Lead response times were slow, follow-ups inconsistent, and there was no way to surface insights from thousands of past conversations. The system needed to learn how Xavier actually communicates — his tone, objection handling, and closing style — and respond to new leads in his voice, 24/7, without hallucinating loan terms or compliance-sensitive information.
How We Solved It
RAG Pipeline & Data Ingestion
Built a full data ingestion pipeline that pulls all GHL contacts, conversations, messages, and call recordings in sequence. Deepgram transcribes call audio; LangChain document loaders chunk the transcripts, SMS threads, and emails into vector embeddings stored in Pinecone — namespaced per sub-account for data segregation.
AI Brain Trained on Xavier's Voice
LangChain RAG chains retrieve the most contextually relevant past conversations when generating each response. Claude serves as the reasoning engine, guided by a carefully engineered style prompt calibrated against Xavier's real messages. Every AI response is scored for confidence — anything below threshold is flagged for human review rather than auto-sent.
GoHighLevel SMS & Email Automation
n8n receives inbound GHL webhooks in real time, forwards the payload to the FastAPI AI server, gets back a Claude-generated response in Xavier's voice, and pushes it through GHL to the lead — all within seconds. Every AI interaction is logged in Supabase and re-vectorized into Pinecone to make the system smarter over time.
Admin Portal with Chat Interface
A Next.js portal built on Supabase Auth gives Xavier and his team a single place to see everything. The chat interface connects via WebSocket to the FastAPI server — Xavier can ask the AI questions in plain English: 'Which leads should I follow up on?' or 'What objections came up most this week?' — and get real answers grounded in his actual data.
Activity Feed & Applications
The Activity Feed shows every AI-generated message with its channel, confidence score, lead snippet, and a manual Send override. The Applications module manages mortgage submissions — draft, submitted, under review — with AI risk scoring and Arrive LOS integration for pushing structured 1003 application data without manual re-entry.
Pattern Analysis & Content Generation
A LangChain agent queries Supabase and Pinecone to surface top objections, successful scripts, rep performance differences, and conversion trends. A content generation module produces social posts, email sequences, and SMS copy — all in True Mortgages' brand voice — on demand from the portal.
Authentication
Dashboard & Leads
Applications & Forms
AI Brain & Activity
Data Pipeline & Settings
Results & Impact
19,712 Active Leads Managed
Full GHL contact base ingested and vectorized — AI handles lead engagement across all active contacts automatically.
90% Average AI Confidence
AI-generated messages consistently hit 90%+ confidence scores against source data before being sent.
< 5s End-to-End Response
From inbound GHL webhook to AI-generated SMS delivered to the lead — under 5 seconds.
Full RAG Knowledge Base
Contacts, conversations, messages, transcriptions, and documents all vectorized into Pinecone — the AI brain improves with every interaction.
Zero Cross-Account Data Leaks
Supabase RLS policies and Pinecone namespace segregation ensure each sub-account's data is completely isolated.
Compliance Guardrails Built In
AI cannot quote specific rates, make approval promises, or trigger Fair Housing violations — hard-blocked at the system prompt level with audit trail on every message.
Technical Architecture
Have a similar project?
I'm open to freelance projects and full-time roles. Let's build something great together.