skip to content
.katana

Loop

Multi-tenant, event-driven order and loyalty platform for restaurants — ingests orders across channels with exactly-once processing, maintains an auditable loyalty ledger, and runs a guardrailed AI retention agent with a published decision-quality eval.

Loop showcase
Role
Full-stack Development, Backend Engineering
Date
September 4, 2026
Deliverables
Event-driven order-ingestion pipeline, Loyalty ledger service, Guardrailed AI retention agent + eval, Storefront audit agent, Live dashboard
Tools
TypeScript, Node.js, Fastify, MongoDB, Redis, BullMQ, Next.js, AWS

Loop simulates a restaurant-ops backend in the shape of a real production system. Orders land through an idempotent outbox pipeline, loyalty balances update transactionally, and an LLM-backed retention agent decides who to re-engage — bounded by send caps, an approval queue, and a full audit log — alongside a storefront-audit agent that reuses the same job and agent infrastructure.

A Fastify/Node.js API writes orders into MongoDB via an outbox transaction, tailed by Change Streams (not polling) for exactly-once delivery with a DLQ and a replay endpoint. BullMQ schedules a retention-agent worker that produces structured, guardrailed decisions — send caps, a human-approval queue, a full reasoning log — against real order history, validated against a hand-written golden-set eval rather than schema conformance alone. A storefront-audit agent reuses the same tenant/queue/decision-log plumbing to score page speed and completeness via Lighthouse. A Next.js dashboard renders a live order/ledger feed over WebSocket and the eval scorecard, deployed on AWS.

view repo

Overview

Loop simulates a restaurant-ops backend in the shape of a real production system. Orders land through an idempotent outbox pipeline, loyalty balances update transactionally, and an LLM-backed retention agent decides who to re-engage — bounded by send caps, an approval queue, and a full audit log — alongside a storefront-audit agent that reuses the same job and agent infrastructure.

Loop

Approach

A Fastify/Node.js API writes orders into MongoDB via an outbox transaction, tailed by Change Streams (not polling) for exactly-once delivery with a DLQ and a replay endpoint. BullMQ schedules a retention-agent worker that produces structured, guardrailed decisions — send caps, a human-approval queue, a full reasoning log — against real order history, validated against a hand-written golden-set eval rather than schema conformance alone. A storefront-audit agent reuses the same tenant/queue/decision-log plumbing to score page speed and completeness via Lighthouse. A Next.js dashboard renders a live order/ledger feed over WebSocket and the eval scorecard, deployed on AWS.

Features

  • Idempotent multi-channel order ingestion with outbox, DLQ, and replay
  • Auditable, transactionally-correct loyalty ledger
  • Guardrailed AI retention agent: send caps, approval queue, full decision audit log
  • Published golden-set eval scoring the retention agent's decision quality, not just schema conformance
  • Storefront-audit agent reusing the retention agent's job/queue infrastructure
  • Live WebSocket dashboard with an eval scorecard
.see also
view all projects
say hello