cxops-ai

CXOps AI is a production-style Agentic AI customer experience platform that combines autonomous AI workflows, Retrieval-Augmented Generation (RAG), human-in-the-loop approvals, Zendesk integration, durable background execution, and AI observability. It is designed to demonstrate how AI agents can safely operate inside a real customer-support environment—not just generate chatbot responses. The platform can ingest customer support tickets, analyze customer intent and operational context, retrieve trusted knowledge using RAG, select an appropriate support action, determine whether that action can execute autonomously, require human approval for higher-risk actions, execute approved actions against Zendesk, record complete audit trails, and monitor AI latency, cost, reliability, and workflow outcomes. The system deliberately separates LLM reasoning from operational authorization. The model can recommend an action, but deterministic application policies decide whether that action is allowed to execute. Core capabilities include: - Agentic AI workflows that produce structured decisions such as respond, route, escalate, internal_note, human_review, and no_action. - Risk-aware tool authorization with policies independent of the language model. Low-risk operations execute automatically; higher-risk actions require explicit human approval. - A complete RAG pipeline for grounding support decisions in trusted knowledge, supporting PDF, TXT, Markdown, and manually created documents. Includes vector similarity search, configurable thresholds, adaptive source selection, citations, duplicate-document protection, prompt-injection protection, and grounded refusal behavior. - Zendesk integration using OAuth and webhook-based ticket events, including ticket retrieval, creation, updates, comments, synchronization, customer lookup, webhook signature verification, idempotent event handling, autonomous internal-note execution, and approval-based external actions. - Human-in-the-loop approval queue allowing operators to review AI decisions, reasoning, RAG evidence, proposed responses, planned tools, and add reviewer notes before approving or rejecting execution. - Autonomous execution for low-risk actions, validated end-to-end against the production Zendesk integration. - Durable background jobs with pending/processing/completed/failed states, retries, configurable maximum attempts, exponential backoff, deduplication keys, and error recording. Failed jobs are retained for auditability. - Safety boundaries preventing local demo tickets from reaching Zendesk unless all conditions are met. Blocked attempts are recorded as audit events. - AI observability recording request telemetry including feature, model, latency, token counts, estimated cost, success/failure, and timestamps. - Agent KPIs tracking total runs, autonomous execution rate, success rates, approval rates, action distribution, risk distribution, and workflow status. - ROI measurement estimating minutes saved, labor value, AI cost, and net value, with a deliberate avoidance of reporting formal ROI until a minimum sample size is reached. - Evaluation workflows for both RAG and agent behavior. Current suites report 6/6 RAG scenarios passing and 12/12 agent scenarios passing, with 60/60 repeated runs passing. The production architecture uses Vercel for the Next.js frontend, Render for the FastAPI backend and background worker, Nhost for PostgreSQL with pgvector, OpenAI for LLM and embeddings, and Zendesk for CX integration. The frontend provides operational workspaces for Home, Tickets, New Test Ticket, AI Agent, Approval Queue, Knowledge/RAG, Runs/Audit Trail, and Observability. The technology stack includes Python 3.11, FastAPI, Pydantic, SQLAlchemy, Alembic, asyncpg, Prometheus metrics, PostgreSQL, pgvector, Next.js, TypeScript, React, Tailwind CSS, GSAP, Lenis, Lucide React, Docker, and GitHub Actions. The project includes local setup instructions, testing commands, and production validation details. It is an actively developed portfolio project with planned improvements including mobile responsiveness, broader evaluation datasets, additional CRM/ticketing integrations, richer tracing, and expanded analytics.