How I Built an AI Arbitrage Engine That Scans 14,000 Products Daily

The Problem That Started Everything




I was doing retail arbitrage manually. Checking Walmart, cross-referencing Amazon prices, estimating FBA fees in a spreadsheet. It worked — barely. Three hours a day to find two or three viable deals. At that rate, scaling meant hiring people who would do the same tedious work.




I'm a software engineer. There had to be a better way.




The First Version




The first version was 40 lines of Python. It hit Walmart's product API, checked the Amazon price via scraping, ran a quick margin calculation, and emailed me if the profit was over $5.




It was terrible. The scraper broke every other day. Amazon blocked my IP. I had no deduplication, so I'd get the same product emailed to me 40 times. But it worked enough to prove the concept.




Building the Real System




The production system — what I now call ArbitrAI — has three layers:




1. The Scanner. Runs on a Mac mini in my home office. It pulls product feeds from Walmart's Seller API, runs initial filtering (price range, category exclusions, basic margin check), and sends candidates to the relay server. Currently scanning around 14,000 products per day.




2. The Relay Server. A Node.js process on a DigitalOcean droplet (137.184.184.27) that receives product webhooks, runs the dual-path routing algorithm, verifies prices in real-time against Amazon and Walmart APIs, calculates true FBA/WFS fees, and makes the channel routing decision.




3. The AI Reviewer. This is where it gets interesting. Before a deal is surfaced to users, it passes through a product reviewer that checks Amazon listing quality, reviews BSR trend data, estimates competition intensity, and applies a scoring model. Deals below threshold get routed to the Alibaba sourcing path instead.




The Dual-Path Router




The routing algorithm is the heart of the system. For each product, it calculates:



• **Amazon path**: FBA fees, referral fees, storage costs, true net margin
• **Walmart path**: WFS fees, category-specific adjustments, verified live price
• **Winner**: whichever channel yields better margin, with minimum thresholds applied



If neither channel passes scoring, the product gets forwarded to an n8n workflow that searches Alibaba for equivalent private-label alternatives. A $14 Walmart toy with 8% margin on Amazon might have a factory equivalent for $2.50 — completely different economics.




The Numbers




Current production stats:

• ~14,000 products scanned daily
• ~340 pass initial margin filters
• ~85 pass AI scoring and reach users
• ~12 get actioned (bought or sourced)



The hit rate isn't the point. The point is I'm looking at 85 curated deals instead of doing 3 hours of manual work to find 3.




What's Next




ArbitrAI is now the engine behind NexArb — the SaaS I built so other sellers can use the same system. The scanner, router, and AI reviewer are all running in production. What I'm building now is the layer on top: better deal surfacing, historical tracking, profitability analytics, and Telegram alerts so you know instantly when a deal hits.




If you're doing retail arbitrage manually today, there's a better way.

Share this article

Related posts