Skip to content
Allan Henslee, PhDResume (PDF)

Case 04Party On Delivery · 2025–2026

One data platform with AI analysts on top

One app and one database for a whole small business, with AI analysts that recommend and a person who approves.

Years
2025 – 2026
Business
Party On Delivery, an Austin event-delivery business
Role
Owner/operator; 2,300+ commits of my own, about 85% of the codebase
§ 01

The problem

A small business ran on a patchwork of SaaS tools, with no single view of orders, customers, marketing and operations.

§ 02

What I built

Starting from the storefront Matt Rundle built, I grew it into one Next.js app that ran the whole business: Stripe checkout, group ordering where each guest pays their share, a warehouse app, an owner dashboard, and the data layer underneath.

  • One Postgres database pulling in Shopify, Stripe, Google Analytics and Search Console, with 30 scheduled jobs (order reconciliation, lead alerts, follow-ups, nightly snapshots).
  • The AI layer. A nightly snapshot feeds detectors that post action cards to one review queue; a weekly briefing email adds an AI-written summary; every shipped change is re-measured 14 days later. On top of that, four Claude Code analyst agents (marketing, operations, finance, SEO) read the same snapshots on demand and recommend what to do next. A person approves anything that changes.
  • MCP server (lets Claude query live business data in plain English): 8 read-only tools, two access levels, rate limits, every request logged.
  • An AI receiving desk. Photograph a supplier invoice and Claude pulls out SKUs, case counts and costs. The warehouse assistant can look things up through 9 tools but can only propose changes; a person approves.
§ 03

My part, and who else

2,300+ commits of my own, about 85% of the codebase. Matt Rundle built the original site, design system, first move off Shopify, early group orders and the first AI chat persona; a partner built the partner-referral hub.

§ 04

How it works

  1. 01Shopify, Stripe, GA, Search Console
  2. 02One Postgres database30 scheduled jobs
  3. 03Nightly snapshot
  4. 04Detectors and analyst agentsRecommend only
  5. 05Review queueA person approves
  6. 0614-day re-checkEvery shipped change
Fig. 1 Recommendations flow to one review queue; every approved change is re-measured two weeks later.

How I know it works

Tests, review, measured results

CheckHowResult

Automated tests

Test files run in CI on every change

217

Pass

A recommendation I rejected

The AI analyst said to drop two referral partners as money-losers. Our tracking under-counted their sales, so I rejected it and added a rule: no partner decisions on that number until the cost data covers 70% of sales.

Rejected

§ 05

Decisions

  • Detectors and agents recommend; only a person approves a change.
  • Read-only tools for the AI, with separate access levels and a log of every request.

Why it matters in healthcare

Read-only AI tools, separate access levels and a log of every request is the access-control and audit-trail pattern any patient-data system needs; the 14-day re-check is post-market surveillance for recommendations.

§ 06

What changed

One view of the whole business, and weekly reporting went from manual pulls to an AI-written email every Monday morning.

Under the hood

109

tables

327

API routes

30

scheduled jobs

217

test files

  • Next.js
  • Postgres
  • Prisma
  • Stripe
  • Shopify
  • MCP SDK
  • Anthropic SDK
  • Vercel