Project Caravel
An AI-operated commerce platform running a global fashion business — built as the first engineer.
Overview
Joined as the first engineer and built the technical foundation for a global commerce and distribution platform. The system consolidates retail and wholesale channels into a single source of truth: a full admin ERP (orders, invoices, purchase orders, shipments, returns, warehouse picking), containerized sync workers pulling products, orders, and inventory from 6+ external marketplaces, and an AI layer that computes SKU-level P&L and drafts decisions for humans to approve.
Presented under a working title — the platform runs a real company's global supply chain, so identifying details are withheld.
Full stack
- Next.js 16 · React 19 · TypeScript 5 · Node 24
- PostgreSQL (multi-tenant) · MySQL legacy integration
- AWS S3 + CloudFront · SQS + Lambda · SES
- Gemini + OpenAI multi-provider abstraction
- Docker · CodeBuild/ECR/CodeDeploy · nginx
- AES-256-GCM field encryption · role-based access
Engineering highlights
The parts a code review would find interesting.
01A job queue with no queue service
Core job orchestration runs on a Postgres-backed distributed queue using SELECT … FOR UPDATE SKIP LOCKED, consumed by separately containerized sync and analysis workers — no external broker to operate, monitor, or fail.
02PII encrypted at the application layer
Buyer PII is encrypted at rest with AES-256-GCM, versioned ciphertext prefixes, and key fallback chains — field- and blob-level — with a documented compliance evidence guide.
03Money math in SQL
The profitability engine is window-function-heavy SQL: lateral joins against landed-cost snapshots, proportional commission allocation across order lines, and rolling 30-day per-channel rollups — with self-test scripts covering the critical financial paths.
04AI agents with guardrails
A reorder agent predicts per-SKU stockout dates and drafts purchase orders gated behind human approval with margin guardrails; generation tooling writes channel-specific copy and AI fashion-model imagery with identity-preserving anchors. GenAI content workflows cut operational overhead by ~90%.
05One TypeScript codebase, every layer
Public site, admin ERP, cron orchestration, containerized workers, and Lambda image-mirroring (SQS → S3/CloudFront URL rewriting) — built and operated end to end, deployed via CodeBuild/ECR/CodeDeploy with 107 incremental SQL migrations and a custom runner.