Skip to content
Plate 01

CreditPulse AI

A voice agent that pre-qualifies mortgage applicants in a single conversation.

Caller voice

Problem

Initial mortgage screening is usually a long form or a phone call to collect the same data every time: income, down payment, property value and debts. It's slow for the applicant and expensive for the lender.

Solution

A voice agent runs the conversation, understands mid-sentence corrections ("actually, with my spouse's income…") and fills a structured dossier. A live dashboard shows debt-to-income, loan-to-value and amortization while the person talks.

Decisions

  1. 01The LLM computes nothing: numbers come from a deterministic Python engine, and the model only receives results through tool calls.
  2. 02Real-time audio pipeline: 16-bit PCM at 24 kHz over WebSocket, with barge-in so the user can interrupt the agent.
  3. 03Indicative statuses driven by a configurable policy instead of a blunt approved/rejected.

Stack

  • Python
  • FastAPI
  • WebSocket
  • AssemblyAI
  • React 19
  • TypeScript
  • Tailwind 4
  • Web Audio API
← Back to projects