Project
gcex-pilot-16862 · no. 804472053350
A customer-support AI agent for a digital lender — it answers borrowers' loan questions, acts on their account, stays compliant, and helps human agents. Built from code on Google's Gemini Enterprise for Customer Experience, with the official CXAS SCRAPI toolkit.
A borrower chats (or talks) to Tilicho Credit Assist about their loan. The demo signs in as one borrower — Asha R., loan TL-1001, verified by loan ID + last-4 phone. Only the loan back-office is a stand-in; everything else is live Google Cloud, and all data is synthetic.
| Looks up loan accounts | Tool calls to the lender's LMS — getAccountSummary, getPayoffQuote — after verifying the borrower (loan ID + last-4 phone). |
| Raises requests | Files tickets for restructuring, complaints or KYC via the LMS createTicket tool, and returns a ticket ID + SLA. |
| Answers policy questions | Pulls from the policy documents & terms-and-conditions stored in the Vertex AI RAG corpus, and quotes the exact figures — not world knowledge. |
| Stays compliant | A guardrail blocks legal / tax / investment advice and guaranteed outcomes, replacing them with a safe reply. |
| Speaks & listens (in chat) | Inside the web chat: 🔊 voice out via Google Chirp-3 HD TTS and 🎤 speech-to-text in via the browser (Chrome/Edge) — a feature of the chat, not a separate telephony channel. CES session memory carries context across turns. |
| Turns chats into insights | Every conversation auto-ingests into the native Contact Center AI Insights product (real Google ML sentiment + entities); the dashboard surfaces that analysis alongside keyword call-driver breakdowns. |
The borrower-facing chat: real tool calls, grounded policy answers, compliance, ₹ formatting, transparency tags — plus 🎤 speak / 🔊 listen in-browser (Google Chirp-3 HD voice) as a feature of the chat.
Every conversation auto-ingests into Contact Center AI Insights (real Google ML sentiment + entities); the dashboard surfaces it + keyword call-drivers.
from cxas_scrapi import Apps, Agents, Tools, Guardrails t = Tools(APP_NAME) t.create_tool("servicing", "Servicing API", tool_type="open_api_toolset", payload=…) # REST tools t.create_tool("policy-rag", "policy_kb", tool_type="file_search_tool", payload=…) # RAG Guardrails(APP_NAME).create_guardrail("compliance", …) Agents(APP_NAME).create_agent("Credit Assist", agent_id="credit-assist", instruction=FULL)
One control plane. Every borrower turn flows through the Google-managed CES runtime — Understand → Ground → Act → Guard → Respond — calling our tools (actions) and Vertex RAG (grounding), then tapping every turn into Contact Center AI Insights. Click a stage to see what it does. The build plane (CXAS SCRAPI) configures it all once.