What does it take to build an Amazon?
I've spent nearly two decades leading enterprise technology programs — so I know how many specialists a marketplace is supposed to require. I wanted to test that. So as an AI experiment, I built a working B2B marketplace system end to end: buyer and vendor AI, payments, logistics, and a self-running content engine — all functioning. It took 17+ distinct roles, every one my own.
Everyone treats building a marketplace as a job for a company — engineers, ops, designers, marketers, each in their lane. Is that still true in the age of AI, or just how it's always been done? The only honest way to find out was to build the whole thing myself and count the hats. The answer turned out to be seventeen.
A marketplace normally spreads this work across a whole company — a different specialist for every box below. I did all of them myself. None of it came from a course; each skill was picked up on the spot, because the build simply wouldn't ship without it. Here are the 17+ hats, grouped by function.
// 17+ hats · 1 operator · 1 question, answered
A list of roles is easy to claim — so here's the proof. Seven working systems sit behind it, each one built, deployed, and validated in production. What each does, and the stack it runs on:
The Autonomous Content Engine
An agent that processes the catalog in batches, generates regional-language search content with an LLM, and applies it through a staging → apply → rollback workflow with backups and per-batch locks. Runs on a nightly schedule, unattended. ~9,300 products optimized to 99% coverage.
The Conversational Assistant
A storefront AI assistant answering buyer questions in plain language. The interesting part was the platform fighting me — the commerce engine strips JavaScript from page content, so I routed it through a server-level injection layer and standalone endpoints to make it work at all. The constraint taught me more than the feature did.
The Buyer Intelligence Tool
A location-aware recommendation engine: a buyer enters their area, an LLM-backed system surfaces the parts most relevant to that regional market. Turns a sprawling catalog from overwhelming into personal.
The Vendor Intelligence Dashboard
The one I'm proudest of visually — a demand-intelligence dashboard styled like a financial terminal, live heatmaps and hot/dead-stock lists across 15 regional markets. Two layers: instant rule-based data on every interaction, and a frontier LLM generating a personalized demand report on request. Built to turn browsers into registered vendors.
The Product Imagery Studio
A bulk image-generation studio for a problem every marketplace has: vendors supply inconsistent, low-quality photos, and hand-editing thousands of SKUs alone is impossible. Pick a product, it generates a full set of standardized marketing shots in one batch and attaches them automatically. Best part: it runs the same prompt through two competing image models side by side, with live cost tracking — pick the better, cheaper result per job. Over a thousand SKUs styled this way.
The Commerce Integration Layer
The plumbing that makes a marketplace real, not a demo: payments, shipping and logistics, transactional email, automated order and payment messaging, and government-API-backed vendor verification for onboarding. Half a dozen third-party systems, wired to behave as one.
The Infrastructure & The Discipline
Linux administration, a CDN/security/performance layer, automated nightly backups — and the part that matters: a written set of operational safety rules. Diagnose with data before acting. Back up affected rows before any destructive change. Make every batch operation safely re-runnable. Anyone can build a system that works once; making one that fails safely is the engineering.
How the whole thing fits together — six layers, from what the buyer sees in their browser all the way down to the servers and outside services it runs on. Every layer below was designed, built, and wired by one person.
The lessons that stuck came from the incidents, not the launches. A run of pages vanished — the lazy move is a blind rollback, but diagnosing first proved the cause predated my work and let me recover cleanly instead of compounding it. A disk-capacity failure silently corrupted tens of thousands of records; fixing that taught me the failure modes that matter are never the ones the demo shows you. Those incidents are why the safety rules exist. The rules are the real product.
One experiment done. Three domains in the lab.
This marketplace was an AI experiment — evidence that one person, with the right AI leverage, can do what used to take a team. Fin-tech, insurance, and travel-tech are next. If you're putting AI into real operations, I'm happy to compare notes.
← Back to all domains