# AI Customer Service

> A support demo connecting cited AI answers, policy-checked tools, human handoff, an agent CRM, and a versioned knowledge CMS.

From a useful answer to a connected human handoff.

- URL: https://weaveward.com/work/ai-customer-service-platform
- Stage: Prototype
- Year: 2026
- Industry: Retail and food delivery
- Context: Internal product
- Capabilities: AI solutions, Web apps, Business systems
- Technologies: React, TypeScript, Vite, FastAPI, Python, SQLAlchemy, SQLite, WebSockets, Google Gemini, FastEmbed, NumPy, BM25 retrieval

## The challenge

A support assistant needs more than fluent replies. It needs the current policy, the relevant order facts, permission to use a tool, and a clear path to a person when judgment or approval is required.

This internal demo explores that connected workflow through fictional shopping and food-delivery scenarios. It brings the customer conversation, the human agent workspace, and the knowledge behind an answer into the same system.

## What we built

The implemented demonstration connects four working areas:

- **Customer experience.** Fictional marketplace and food-delivery catalogs, orders, status updates, complaints, and a help conversation provide realistic workflow context. Commerce and payments are simulated
- **Grounded AI support.** The assistant uses retrieved, approved knowledge and structured order information to prepare cited answers. A deterministic offline fallback supports the demo when the AI provider is unavailable
- **Policy-checked tools.** A backend executor checks ownership, policy, risk, limits, confirmations, and idempotency before recording a simulated refund, cancellation, or other action. The model cannot directly change the database
- **Human handoff and CRM.** Staff can work from a queue, take over a conversation, reply, review suggested answers, and route actions requiring supervisor approval
- **Versioned knowledge CMS.** Documents pass through ingestion, extraction, chunking, review, approval, and publishing. Retrieval uses the versions appropriate to the conversation's access context
- **Operational activity.** Metrics and audit views make conversation, tool, and retrieval activity visible without pretending those views establish a business improvement

Realtime connections keep customer and staff workspaces aligned as a conversation moves between the assistant and a human agent.

## How it works

### React interfaces, FastAPI backend

React and TypeScript provide the customer and staff interfaces. A Python FastAPI backend owns the source records, permissions, domain rules, and action executor.

### A guarded conversation runtime

The conversation runtime combines structured understanding, knowledge retrieval, workflow tools, and response checks. Hybrid retrieval uses vector and keyword signals. A response guard prevents the assistant from claiming an action succeeded without a confirmed tool result.

### Local data, documented deployment paths

The current demo uses SQLAlchemy and a local SQLite database, with a NumPy-based vector index and FastEmbed. PostgreSQL, pgvector, and Redis have documented deployment paths; no production-scale distributed deployment is claimed.

## Current stage

This is an implemented internal demonstration with synthetic data and simulated commerce actions. It does not move live customer funds and is not an integration with a named marketplace or payment provider.

The repository contains backend policy, security, retrieval, handoff, and scenario tests, plus frontend and browser-test tooling. Those suites were not rerun during this portfolio update. Gemini and scanned-document OCR require configured provider access for separate live validation.

No real support client, public production deployment, resolution rate, customer-satisfaction improvement, or cost reduction is claimed.

## Where this fits

Support becomes more useful when answers, records, and human decisions connect. We build [AI solutions](/services/ai-solutions) around your knowledge and workflows, and pair them with [business systems](/services/business-systems) and [web applications](/services/web-apps) that help your team act. [Discuss your support workflow](/contact) with Weaveward.