All case studies

04 · fintech · back office

Laravel Nova AI copilot: streaming chat and grounded reply drafts

Staff ask about a loan or draft a customer reply without leaving the admin panel, and every draft waits for a person to review and send it.

  • Laravel Nova
  • Vue 3
  • Node.js
  • TypeScript
  • LangGraph
  • WebSockets
  • Redis pub/sub
  • Hono
Client
Stepping Stone Loans, a US consumer lender
Role
Senior full-stack & AI engineer
Period
Dec 2024 – present
Engagement
Long-term contract through Upwork

Problem

Support staff switched between the admin panel, the contact-center tool and documents to answer customers, and early AI drafting lived inside the contact-center plugin.

Answers had to stream into every open tab as they were generated instead of arriving all at once.

How it works

diagram · How a reply is drafted
  1. 01Open customer contextStaff open the copilot with a loan or customer attached.
  2. 02Stream a draftA Node.js agent gathers approved knowledge and account facts; the draft streams into every open tab over WebSockets.
  3. 03Review and editStaff copy, edit or reject the suggestion.
  4. 04A person sendsA person sends the reply. Nothing is sent automatically.

What I built

  • Built a custom Laravel Nova tool in Vue 3: streaming chat, loan search and context, a draft-reply mode and a mode for proposing knowledge-base changes.
  • Built grounded reply drafting: drafts use approved knowledge and the customer’s account facts, and personal data is redacted before any model call.
  • Wrote a TypeScript Node.js service that takes jobs from Laravel through a Redis queue and streams answers back over a WebSocket server with heartbeats, multi-tab sessions and graceful shutdown.
  • Built a streaming chat client that reconnects with backoff and keeps every open tab in sync.
  • Added an LLM health dashboard in Nova: cost per draft, latency by model version, evidence coverage and how much staff edit the drafts.

Key decisions

Grounded or nothing
Every figure in a draft must come from real account data, or the assistant declines to draft.
Stream, then fan out
Answers stream as they are generated, and Redis pub/sub fan-out is designed for several worker servers, so every open tab gets each chunk.
One AI panel for the back office
Reply drafting moved out of the contact-center plugin into this panel, with tests that guard the removal.

Guardrails

  • The assistant drafts; staff decide and send.
  • Role-gated: support agents see a reduced mode.

Testing and evaluation

  • Pull requests that touch an AI path replay a reviewed evaluation set in CI before merge.
  • Component tests cover the chat and drafting modes; Node.js unit tests cover transport, concurrency and rate-limit retries.

Result

One place for AI help across the back office, with cost and quality visible by model version.

Answers reach staff as they are generated, in every open tab.

media note

The lender’s internal systems are private, so this case study uses words and a diagram instead of screenshots.