A mortgage brokerage needed to re-engage past leads at scale while eliminating hours of daily manual document review. Their small team was stretched thin — unable to keep up with the volume of outbound follow-ups required to stay competitive, and losing productive hours each day to the repetitive validation of identity documents, bank statements, and proof of income. Leads were slipping through the cracks, and document backlogs were creating bottlenecks that delayed the entire origination pipeline.
We designed, built, and deployed 10 specialized AI voice agents purpose-built for outbound remarketing calls. Each agent was connected to Retell AI and VAPI for telephony, enabling high-quality phone conversations at scale without human intervention.
During live calls, agents perform real-time tool calls — checking calendar availability, booking meetings directly into the broker's schedule, and gathering situational data from the lead to qualify interest and urgency. Every call feeds into a post-call analysis pipeline that scores sentiment, extracts key information, and automatically schedules follow-ups for the human team.
The biggest technical challenge was hallucination control. Early in testing, agents would go off-script, fabricate details, or drift from the intended conversation flow. We solved this through an intensive two-week sprint of systematic prompt engineering, guardrail implementation, and adversarial testing — producing agents that stay on-message while still sounding natural and responsive.
Parallel outbound calling with independent conversation state and per-agent specialization.
Live availability checks and meeting booking during calls — no human handoff required.
Automated sentiment scoring, information extraction, and intelligent follow-up scheduling.
Conversation boundaries and prompt safeguards that keep agents on-script without sacrificing natural delivery.
End-to-end automation from lead list ingestion through outbound calls to CRM updates and follow-up scheduling.
Alongside the voice agents, we built an automated document extraction and validation system to eliminate the manual review bottleneck. The system processes bank statements, proof of income, and identification documents using a combination of Mistral OCR and OpenAI Vision models — handling multi-format inputs including scanned PDFs, photographs, and digital exports.
Documents are routed through processing queues with built-in retry logic and rate limit management, ensuring reliable throughput even during peak volume. Extracted data is validated against configurable rule sets before being passed downstream.
A standout feature is the self-serve extraction prompt system: end users can define and add custom extraction rules for new document types without engineering involvement. This means the system scales with the business — as new document requirements emerge, the team configures them directly rather than waiting on development cycles.
Whether you need voice agents, document automation, or both — we can design a system tailored to your workflow and scale.
Take the Quantum Leap