Stella D.

I build production AI customer support agents for Shopify and DTC brands using n8n, Claude, and OpenAI. Most stores can auto-resolve 60 to 75 percent of inbound tickets the moment the right system is in place — WISMO, returns, restock requests, routine product questions — and the agents I build do that without ever inventing an order number, tracking link, or refund commitment. What makes them production-grade (not chatbots): every agent follows the same proven pattern. Webhook to small-model classifier to parallel tool calls (live Shopify lookup, KB retrieval, refund guard) to a brand-voice drafter (Claude Sonnet) to an independent confidence judge gating auto-sends at 0.85, then either auto-send or Slack escalation with Approve, Reject, and Edit buttons. The agent never invents order numbers, tracking links, restock dates, or refund commitments — it quotes from your real Shopify data and your grounded KB only. Anything risky (refunds over a set amount, address changes after fulfillment, lawyer or chargeback keywords, VIP-tagged customers) always escalates to a human. Before going live, I run dry-run mode against your last 200 plus historical tickets and share a full accuracy report so you see exactly which intents to trust day one. After launch, you get a daily Slack digest covering volume, auto-resolve rate, top intents, and escalations cleared, plus documentation your team can use to update prompts and the KB themselves. No lock-in. Stack: n8n (cloud or self-hosted on Hetzner / DigitalOcean), Anthropic Claude (Sonnet 4.5, Haiku), OpenAI (GPT-4o, GPT-4o-mini), Shopify Admin API, Slack, Supabase pgvector, Gorgias and Zendesk webhooks, Klaviyo, JavaScript, Python. Typical engagement: 5,000 to 15,000 dollar fixed-price builds, 2 to 4 weeks from kickoff to live. Optional 750 dollar per month retainer covers KB updates, prompt tuning, and weekly accuracy reviews. Cancel any time after 90 days, no lock-in. Available US Eastern hours. If your support inbox is falling behind and you want an agent that sounds like your brand and never makes things up, send me a message. I'll look at your FAQ and one recent ticket exchange before quoting, so the proposal is specific to your store, not a template.