EnterpriseLiveSole engineer — built & shipped end-to-end
Craig

AI agent that quotes, charges and manages print-shop orders.

Overview

Craig serves a Dublin print shop's customers over web chat and email: it gathers specs through interactive UI (chips, sliders, to-scale size cards), prices jobs exactly from the live catalogue — never guessing — and runs the full cycle: approval, payment links, production tracking on a kanban board, artwork reminder sequences and post-delivery reviews. It turns inquiries into paid orders with zero human touch for standard cases.

Highlights

  • An LLM that never invents prices: strict tool-calling against a deterministic pricing engine (6 strategies) plus ~15 server-side anti-hallucination gates
  • Spec machine generated from the live catalogue: quantity sliders, size cards drawn to real scale, multi-step flows where impossible options simply disappear
  • Full order funnel: payment webhook → drag-and-drop production kanban → automated artwork reminders → post-delivery reviews with HMAC-signed tokens
  • Multi-channel, one mind: the same agent runs the web widget and the email inbox (confidence-based triage and threading), multi-tenant from day one

Architecture

  • Embeddable vanilla-JS widget (zero build) plus a Next.js admin dashboard for approvals, kanban, catalogue and attribution analytics
  • FastAPI backend with a pure pricing engine decoupled from the LLM, idempotent migrations, scheduled jobs and payment webhooks; multi-tenant Postgres
  • LLM with strict function-calling (the model can only call pricing/escalation tools), an email triage classifier, and a gate layer validating every reply

Tech stack

PythonFastAPIDeepSeek (tool-calling)PostgreSQLNext.jsStripe ConnectCloud RunSupabaseTailwind

Media

Video placeholder
Image placeholder