Chris Pallé
↑Overview

01 · AeroTorch · Manufacturing · 2025–2026

A torch maker, selling online for the first time.

Problem
Product knowledge lived in a paper catalog and the owner’s head. No way to sell direct.
Action
One product database first, then a store, search and feeds on top, built with AI agents inside checks I set.
Solution
A site and database as the cornerstone of its digital transformation. First sale the day the store opened.
  • 4 channelsOne catalog feeds the site, the store, a distributor’s import, and Google and AI shopping feeds
  • 423 specsAnd 356 links across 70 products, modeled once
  • 7 weeksFiling cabinet to working catalog
Exploded view: product records at the bottom, admin in the middle, and on top the four channels the product data reaches: the storefront and search today, distributors and supply-chain vendors by API when built. One product runs up through all of it.
Role
Product, UX, data model, build, store and go-to-market
Stack
Next.js · Supabase · Stripe · Clerk · Vercel · SendGrid · PostHog
AI, in order
ChatGPT → Replit agent → Claude Code with checks
The full story · 5 chapters

Overview

Main street, built like a platform.

AeroTorch makes torches. The owner can tell you without looking which tip fits which handle and what goes in every kit. Almost none of it was written down: a printed catalog, a stack of manuals, and a folder of photos still named by the camera.

Today the store sells to strangers. Behind it is a 70-product catalog the business owns outright, with every spec, manual and part relationship in one place. The same records feed the website, the store, a distributor’s import file and the shopping feeds that Google and AI assistants read.

The full story, chapter by chapter

I · What changed for the business

The most important record in the business was the owner’s memory.

You can now ask the catalog a question nobody could answer before without the owner in the room: which kits use the 2AM handle? Twenty of them.

AreaBeforeNow
Product knowledgeDecades of it, held by the owner and a printed catalog.A database: 70 products, 423 specs, 356 links between parts and kits, every manual attached.
Selling directNo way to take an order online.A store with card checkout, tax, shipping rules and refunds. First sale the day it went live, September 11, 2026.
Working with distributorsRetyping products into each partner’s format.One export with every product, spec and relationship, ready to import.
Being foundGoogle had indexed 4 pages and not one product.Every product page readable by search, plus feeds to Google Shopping and AI shopping.
Keeping it currentBy hand.Automated checks every Monday: site health, listings, dealer prices and mentions.
What the owner doesAnswers every product question himself.Confirms facts, sets stock and shipping rules in one admin screen, ships orders.

II · How it happened

He asked for a website. We started with Aero’s first digital foundation underneath it.

Data first, then one capability at a time as the business was ready. A few hours a day, in bursts, and each burst shipped something real.

Layer 1 · Records

Read every product for what they share.

Seventy products, each written up its own way, came down to 40 shared spec definitions and 14 kinds of link between parts and kits. Every product is now one record built from them.

TIPHANDLETANKFLAMEKIT CONTENTSSKUTIPHANDLETANKAKJ5-SAJ4-T2AM-TB or MC70 PRODUCTS, EACH WRITTEN UP ITS OWN WAY40 SHARED SPEC DEFINITIONS70 RECORDS · 423 SPECS · 356 LINKS

Layer 2 · Admin

The admin is that structure, made editable.

The owner confirms facts, links parts to kits and sets stock and shipping rules in one place. Change it once and every channel has it.

ADMIN · PRODUCTSProductsCategoriesSpecsMediaOrdersAKJ5-S air acetylene JET kitONE RECORD · B OR MC TANKSPECSTipAJ4-THandle2AM-T quick disconnectRegulator440G or 410GFlameJET, high temperatureLINKSAJ4-T tip2AM-T handle440G regulator+ kits that share themSTOCKPHOTO NAMEscore 0–100

Layer 3 · Channels

One record, every channel.

The store, the product pages search engines read, a distributor’s import file and the Google and AI shopping feeds all draw from the same row.

AKJ5-SONE RECORDSTOREFRONTPRODUCT PAGES IN SEARCHDISTRIBUTOR IMPORT FILEGOOGLE AND AI SHOPPING FEEDS<item><g:id>AKJ5-S</g:id></item>
The Duplex winter landing page for copper roofers.
The new air acetylene torch kits category page, with filters and kit cards.
The new AKJ5-S product page: photo, B or MC tank choice, price and shipping details.
The new All products page on a phone.

New pages, in design. They go live with the owner’s sign-off.

III · How AI did the work

The system came first. The AI tools changed three times around it.

The groundwork held while the tools changed: a chat window to read the old catalog, a prototyping agent for the first site, then agents working inside checks I set. Each time I trusted them with more.

  1. 01 · ChatGPT · Aug–Sep 2025

    A chat window

    ~490 messages in 8 days

    Scans of the old catalog turned into tables: names, specs, kit contents.

    What I trusted it with

    Reading. I checked every answer.

  2. 02 · Replit · Aug–Oct 2025

    A prototyping agent

    ~560 commits in 7 weeks

    The first site and catalog fast, with no branches or review. Rolled back eleven times.

    What I trusted it with

    A site to show distributors. Not a store.

  3. 03 · Claude Code · from Mar 2026

    Agents with guardrails

    Day 2 every change checked

    Rules, a spec and skills on day one; lint, tests and a security scan on every merge from day two.

    What I trusted it with

    A store that takes real money, with my review on every change.

IV · How it runs now

It proposes. We decide.

This is the harness. Every week it reads the live business and turns what it finds into proposals: a fix, a new page, the next thing to try. The owner and I pick what happens next from the numbers, and the agents build it inside the checks.

01 · WATCH02 · PROPOSE03 · FIX AND GATEEvery Monday a check readsthe live site, the week’s searchnumbers and six dealers’ pages.Products showing out of stock15 broken linksNo returns pageAgent’s pull requestlint · tests · security scanMy reviewShips04 · LEARNA wrong reading becomes a rule the routine checks from then on.

V · For your team

The groundwork-first pattern a platform team uses, at main-street scale.

Forward-deployed work, at small scale. Plenty of companies run on knowledge that lives in a few people’s heads. Get it into a form the business owns, then build on it in steps people will actually use.

What carries over

  • I

    Get into an unfamiliar business fast

    Learn how it really runs, find where the knowledge lives, and decide what to build first.

  • II

    Set agents up to do real work

    Context, rules, checks and a loop that ships, set up in a day on a new codebase.

  • III

    Stay in the loop where judgment matters

    Decide what the agents decide, what they propose, and what a person approves.