Arventa

Arventa is a multi-tenant platform that automates accounts-receivable follow-up for businesses chasing overdue invoices. It combines AI voice agents that place outbound collection calls with AI-drafted emails, and a visual workflow engine decides who gets contacted, through which channel and when. Finance teams get one place to manage receivables and see every interaction with each debtor. Inside the app, users import invoices from CSV or Xero, design collection workflows in a node-based editor (calls, emails, waits, conditional branches on call or email outcomes, record updates and escalation to a human) and launch them on single invoices or in batches of up to 500. Each invoice gets a unified timeline with call transcripts and recordings, emails and inbound replies, notes and an edit history, plus an AI-generated summary of where the account stands and what should happen next. Calls the agent cannot resolve become follow-up items in an Action Needed queue: a customer asking for a human, wanting to negotiate, a wrong number or an unreachable debtor. Admins can listen to calls live and configure contact hours, the voice agent and verified sending domains for each organization. The system is a monorepo with three deployable parts: a React dashboard, a Node.js call server that also runs the workflow engine, and a Supabase project with 74 SQL migrations and 11 edge functions. It was built by a five-person engineering team over about six months.

Collecting overdue invoices is repetitive, manual work. Collectors call debtors one by one, leave voicemails, send reminder emails, note what was promised and try to remember when to follow up. Context ends up scattered across phone logs, inboxes and spreadsheets, calls placed at the wrong hour or to a wrong number waste effort, and for any given account it is hard to tell what has been tried and what should happen next. Arventa turns that process into durable, auditable workflows. A Postgres-backed state machine on pg-boss runs each workflow: it parks a run while waiting on a phone call, an email reply or a timer, and resumes it when a telephony or voice-agent callback or an inbound email webhook arrives. Each run executes against a frozen snapshot of the workflow, so editing a workflow never affects runs already in flight. Outcomes are classified into structured categories (voicemail, wrong number, promise to pay, escalation and others) that drive retry policies and branching, and calls and emails are deferred until the debtor's local business hours, with the timezone inferred from the phone's area code. The product depends on combining three things in one system: conversational AI calling, per-organization configurable workflows, and outcome-aware follow-up that feeds a human action queue. The code reflects that integration-heavy core: Twilio for telephony and answering-machine detection, ElevenLabs for the conversational agents and their tools, Claude models for classification, summaries and email drafting, and Resend for outbound email and inbound replies.