Back to Portfolio
Featured Project

True Mortgages — AI-Powered Mortgage Platform

AI brain for a US mortgage business that ingests call recordings, SMS threads, and emails to auto-respond to GoHighLevel leads in the owner's communication style. Full admin portal with RAG-powered natural language queries, pattern analysis, content generation, lead management, and Arrive LOS integration for automated 1003 application generation.

True Mortgages — AI-Powered Mortgage Platform

Tech Stack

FastAPINext.jsLangChainAnthropic ClaudePineconeSupabaseDeepgramGoHighLevel APIn8nAWS ECSRedis

Overview

Built a centralized AI platform for Xavier Alcala's mortgage business (XA Financial Services / True Mortgages). The system ingests all existing GoHighLevel call recordings, SMS threads, and email conversations — Deepgram transcribes the calls, LangChain chunks and embeds everything into Pinecone, and the RAG pipeline trains the AI on Xavier's actual communication style. When a lead texts in at 9pm, the AI responds through GHL the way Xavier would — not generic bot text, but his tone, phrasing, and objection-handling style drawn from 15,000+ real conversations. The admin portal gives Xavier a single command center: a natural language chat interface to query the AI brain ('what objections came up most this month?'), a pattern analysis engine surfacing conversion trends and rep performance, a content generation module for social posts and email sequences in brand voice, and full lead and application management. The platform also powers the truemortgages.com website with GHL-integrated lead capture and a templated sub-website system for each team member — giving every rep their own A2P-compliant web presence under a shared brand. Milestone 3 connects Arrive LOS to extract 1003 mortgage application data directly from conversation history.

Key Features

Full GHL data ingestion pipeline: call recordings → Deepgram transcription → LangChain chunking → vector embeddings → Pinecone, covering 15,000+ existing leads
AI SMS and email auto-response via GHL — responses generated in Xavier's communication style using RAG context from real conversations
Natural language AI brain chat: Xavier types 'what objections came up most this month?' and gets ranked insights backed by actual Pinecone-retrieved data
Pattern analysis engine: identifies top objections, conversion scripts, rep performance differences, lead drop-off points across all GHL conversations
Content generation module: social media posts, email sequences, SMS copy — all generated in Xavier's brand voice
Admin portal: dashboard with AI activity feed, response metrics, lead engagement stats, and conversion trends
Multi sub-account support with full data isolation via Supabase Row-Level Security — reps cannot access each other's data
truemortgages.com main website (Next.js SSR, schema markup, SEO) with lead capture forms piped directly into GHL
Templated team member website system: each rep gets their own site (irvingrivera.com, rudysoto.com) — same brand, swappable name/bio/headshot, wildcard DNS routing
Arrive LOS integration (Milestone 3): LangChain extracts 1003 mortgage application fields from conversation history, flags low-confidence fields for human review before submission
AI safety guardrails: topic restriction (no rate quotes or approval promises), hallucination confidence scoring (< 85% confidence → human review queue), human escalation on angry/complex leads, full audit trail of every AI message sent
Deployed on AWS ECS Fargate with CloudFront CDN, GitHub Actions CI/CD, Sentry error tracking, and PostHog analytics