A single AI platform for customer communication: chat, automation, and an AI call center.
FlowOne lets teams build AI agents that talk to their customers on their website, in messaging apps, and over the phone — no programming knowledge required. Those conversations then turn into real work: a CRM record is created, an email is sent, an approval is requested, a follow-up happens. One knowledge base, one inbox, one workspace.
FlowOne is a production-ready, multi-tenant SaaS platform built with React, FastAPI, and PostgreSQL. It combines real-time communication, RAG-powered AI agents, visual workflows, WebSockets, OAuth 2.0, Docker, GitHub Actions, and integrations with OpenAI, Anthropic, and Google AI in a single architecture.
This project was designed and developed by Sude Özübek and Halil İbrahim Kurnaz.
The platform is built with React, FastAPI, PostgreSQL, and Redis on the application side, and integrates OpenAI, Anthropic, Google AI, ElevenLabs, and Twilio. Core engineering concepts include retrieval-augmented generation (RAG), vector search, the Model Context Protocol (MCP), WebSockets, OAuth 2.0, and multi-tenant SaaS architecture. It is deployed with Docker and Cloudflare, with CI/CD on GitHub Actions and end-to-end testing with Playwright.
A company that wants to bring AI into its customer operations today has to buy three separate products: a chatbot for the website, an automation tool for routine work, and a call-center platform for the phone. Each one keeps its own copy of company knowledge and its own fragment of customer history. The chatbot doesn't know what the customer said on the phone yesterday; the automation tool knows nothing about either.
FlowOne merges this stack into one platform:
| Chat Agents | Answer customers on the website, on WhatsApp and other messaging channels, and in embeddable widgets |
| Workflows | Turn conversation into action: send the email, update the CRM, request the approval |
| AI Call Center | Handle inbound and outbound phone calls with natural-sounding voice agents |
The three layers share everything that matters: the knowledge base, customer memory, and the team inbox. A question asked in the web widget and a phone call the next day are one continuous conversation, not two records in two systems.
FlowOne is designed for small and mid-sized businesses and growing teams without technical staff. Setup is guided and conversational: you describe what you need in your own words, and the platform builds the agent or workflow for you. AI absorbs the volume; critical decisions always stay with a human.
The screenshots below are representative sections of the product. They show the highlights, not the full platform, which spans dozens of screens.
Agents are created and managed in Agent Studio. The user first picks an agent type (chat, voice, or 3D avatar), then a creation mode matching their experience level, from fully guided flow to expert configuration. Every agent is tried in a live test area before it faces a customer. A template gallery covers common roles such as customer support, recruiting, and document analysis. Under the hood, agents run on OpenAI, Anthropic, and Google models; every run is traced and cost-attributed.
Chat agents draw on the company's own knowledge base to hold multi-turn conversations in the web widget, on hosted chat pages, and across messaging channels. A short business description is enough to set one up: the AI-assisted builder fills in the personality, tone of voice, and starting content. Responses stream in real time; when a human is needed, the conversation lands in the shared inbox and a teammate takes over exactly where the agent left off.
Agent setup is guided and conversational. The user describes the agent's job in their own words or picks a sample type; "Generate with AI" expands that short brief into a complete agent definition. Throughout setup, the live preview widget on the right reflects every change instantly, so the agent is tested exactly as customers will see it.
Before publishing, the agent's name, greeting, description, and all knowledge sources are reviewed and edited inline on a single screen. Websites, documents, and curated Q&A pairs form the agent's knowledge base; tone preferences and business rules such as "never invent prices" shape its behavior before it ever faces a customer.
Automations are built on a visual node-graph canvas. You can also describe the flow in plain language and let the AI co-designer build it for you. A workflow connects triggers, conditions, AI steps, and third-party actions, and can pause for human approval before a sensitive step. Every run can be traced step by step, with progress streaming to the run screen in real time.
Workflows are also what make the other two layers useful: a chat escalation can trigger an automation, and a finished phone call can create the CRM record automatically.
Building a workflow with the AI co-designer feels less like filling out a form and more like talking to an architect. The user writes their operational goal in their own words ("When a new real-estate inquiry arrives, analyze the customer's budget and preferences with AI and send matching listings…") or picks one of the ready-made examples.
The platform takes it from there: it analyzes the goal, fills in missing business context (target audience, which CRM is used, approval points) through a short guided conversation, and presents the result as an executable workflow ready for review and approval. The user never touches technical detail; "Activate Workforce" takes the flow to production in one click, and every run can be traced step by step.
Voice agents carry the same knowledge and personality to the phone, running on ElevenLabs speech synthesis and Twilio telephony with real-time speech-to-text. The voice studio covers the full lifecycle: persona design, phone number provisioning, embeddable voice widgets, outbound call campaigns, conversation branching, compliance settings, A/B testing, and a go-live checklist. Supervisors can monitor calls in real time, and every call is quality-scored afterwards.
The voice studio takes a voice agent from idea to real phone calls in minutes. The agent's personality and greeting are defined in your own words or filled in with one click using AI. Then a voice is chosen from the ElevenLabs library, filtered by language, accent, gender, and age, with dozens of natural voices that can each be previewed with one click. Advanced users can switch the speech engine or use their own cloned voice.
Before provisioning a real number, the agent is tested in a real-time voice conversation in the browser using microphone access — a faithful rehearsal of the phone experience.
When ready to publish, a phone number is provisioned in the United States, the United Kingdom, or Turkey, and inbound calls are switched on. From the same screen, the voice widget is added to a website with domain verification and an embed code. Persona settings run from one source for both phone and web.
Every call is evaluated automatically: conversations are scored, recurring topics are discovered, and the sources used in answers are tracked. Conversations can be filtered by date, status, language, and duration.
Teams upload documents (PDF, Word, Excel) or connect web pages; content is managed centrally on the platform. Chat, voice, and workflow agents all answer from this single source of truth: update once, and everything is updated.
All channels meet in a single inbox: web chat, messaging apps, call transcripts, workflow approvals, and agent activity — all updated live over WebSockets. Team members watch conversations as they happen, take over from the agent when needed, and resolve approval requests without switching tools.
The integration hub connects agents to the tools businesses already use: WhatsApp, Slack, Telegram, and Discord on the messaging side; HubSpot and other CRMs; Shopify for e-commerce; Google and Microsoft tools for email and calendar. Connections are set up with standard OAuth or API keys. Advanced users can also add external tool servers through the Model Context Protocol (MCP).
Beyond the hub, FlowOne ships a browser automation agent that completes goal-driven tasks on the web, and a Chrome extension that brings FlowOne playbooks into everyday browsing: summarize a page, save a lead, or draft an email without leaving the tab.
FlowOne is multi-tenant from the ground up. Every workspace has isolated data, its own members with role-based permissions, and its own subscription. Plans are self-service, with live usage indicators and consistent limit enforcement across all modules. Multiple payment providers and currencies are supported, and an enterprise mode covers manually managed contracts. Agencies and partners can apply their own branding through white-label configuration.
Analytics closes the loop: conversation volumes, workflow outcomes, call quality scores, and AI usage costs are visible per workspace. A separate admin panel covers platform-wide oversight and audit logs.
An assistant embedded in the product follows the user step by step. It guides in product context, answers questions, and — with confirmation — performs actions directly. Most users get from sign-up to their first working agent without ever opening documentation.
FlowOne is a multi-tenant SaaS platform developed end to end. The stack is what you would find on any company's engineering page:
| Layer | Technologies and concepts |
|---|---|
| Frontend | React, Vite, Tailwind CSS, an interactive node-graph canvas for workflow editing, WebGL/3D rendering for avatar agents |
| Backend | Python, FastAPI; async-first REST and WebSocket APIs; distributed background workers for long-running jobs such as call campaigns, document processing, and scheduled tasks |
| Data | PostgreSQL, distributed caching and session management, enterprise-scale record keeping |
| AI | OpenAI, Anthropic, and Google models; knowledge-base-grounded answer generation; task-oriented agent orchestration; tool use via the Model Context Protocol (MCP) |
| Voice | ElevenLabs speech synthesis and voice cloning, Twilio telephony, real-time speech-to-text |
| Real-time | WebSocket and server-sent events infrastructure powering live chat, streaming agent runs, and live call supervision |
| Security & compliance | OAuth 2.0 integration flows, role-based access control, encrypted credential storage, PII detection and masking, configurable guardrails, audit logs |
| Observability | End-to-end LLM tracing with per-request token and cost attribution, error and performance tracking, structured logging |
| Deployment | Docker, Cloudflare edge network, GitHub Actions CI/CD; automated unit, integration, and end-to-end (Playwright) test suites |
A few notes for a technical reader, without going into implementation detail:
- One platform, not three. Chat, workflows, and voice are not three products lined up behind the same login. Identity, knowledge, memory, billing, and the inbox are shared at the platform level. That was the hard part — and it was built in from the start, not bolted on.
- Multi-tenant architecture at the core. Workspace isolation, per-tenant roles, and plan limits work the same way in every module, from agent runs to call records.
- No single AI vendor lock-in. The platform runs interchangeably on OpenAI, Anthropic, and Google models. Every AI call is cost-attributed and traceable.
- Real-time was not an afterthought. Live conversation streams, instant visibility into running workflows, and live call supervision are the product's default experience.
- Privacy is taken seriously. Personal data is automatically detected and masked; KVKK (Turkish data protection law) requirements are supported from day one alongside international norms.
- A human is always in the loop. Approval, takeover, and escalation are not decorative extras; they are core concepts every module is built on.
- Not just developed — operated. The platform runs in production behind CI/CD. Tracing, monitoring, and cost controls were planned from day one.
- Bilingual. The application ships in English and Turkish, with right-to-left script support on the marketing side.
FlowOne was designed and built entirely by a team of two. The sections below summarize the engineering responsibilities we personally carried on the platform:
Application layer
- Frontend architecture based on React, Vite, and Tailwind
- Interactive node-graph editor and WebGL/3D integrations
- Async-first backend services with Python and FastAPI: REST and WebSocket APIs, plus background services for document processing, call campaigns, and other long-running work
- PostgreSQL database design: multi-tenant data model, migration management, enterprise-scale record keeping
Platform and infrastructure
- Multi-tenant architecture: workspace isolation, role-based access control, and consistent plan-limit enforcement across all modules
- Authentication and OAuth 2.0 integration flows, encrypted credential storage
- Containerized deployment with Docker and secure publishing through Cloudflare Tunnel
- CI/CD with GitHub Actions: automated unit, integration, and end-to-end test pipelines
- Real-time communication layer with WebSockets and server-sent events: live chat, streaming agent runs, live call supervision
AI and voice
- OpenAI, Anthropic, and Google AI integrations: a single interface that can switch between providers, with per-request token and cost attribution
- Voice infrastructure on Twilio and ElevenLabs: phone number provisioning, real-time speech-to-text, embeddable voice widgets
Business and operations
- Billing and subscription system: multiple payment providers, multi-currency, self-service plans, and an enterprise contract mode
- Tracing and observability: end-to-end LLM tracing, error and performance tracking, structured logging
- End-to-end test suites with Playwright and continuous operation of the production environment
- Deeper CRM and e-commerce integrations
- Scheduled and event-based workflow triggers
- Expanded team collaboration and permissions
- Enterprise capabilities: SSO, organization hierarchies, audit log export
- Marketplace growth for community-built agents and templates
- Multi-region phone coverage
- Public API and developer documentation
FlowOne is an independently developed, production-grade SaaS platform, currently in early access with a live production environment. Its engineering surface spans a rich web application of dozens of product screens; a backend covering agents, workflows, voice, billing, and multi-tenancy; real-time infrastructure; a Chrome extension; and end-to-end deployment automation. All of it was designed and built as a single, coherent product.
This repository was prepared to present FlowOne's product and engineering approach.
The source code is not included here. The materials shared serve as a product showcase, illustrating the platform's scope and technical surface.
Copyright © 2026 Sude Özübek & Halil İbrahim Kurnaz. All rights reserved.
All content in this repository (text, images, and product showcase materials) is protected by copyright. It may not be copied, reproduced, or used commercially without written permission.