Plain Freight, the AI-run China to USA freight agency I founded

Most of my work is software: backend systems, cloud infrastructure and AI in production. Plain Freight is a physical business, and that is the point. Freight forwarding for small shipments is slow, manual and full of repetitive coordination: quoting, chasing a warehouse, filling customs forms, copying tracking numbers between systems, answering the same three questions for the hundredth time. Almost none of that needs human judgement. All of it needs someone who never forgets a step.

Big forwarders are built for big accounts. Ship containers every week and you get an account manager and good rates. Ship 100 g to 2,000 kg at a time, which is the reality for most Amazon FBA sellers, Shopify brands and independent importers, and you are too small to matter. You overpay a courier, or you fight a portal that assumes you already speak HS codes, incoterms and customs bonds.

Then duty-free de minimis treatment was suspended for goods from every country, so shipments valued at $800 or less stopped being exempt and became subject to all applicable duties, taxes and fees (US Customs and Border Protection fact sheet). A wall of customs paperwork and duties landed on exactly the people with the least support to handle it. Plain Freight exists to take that wall away and quote one honest number instead. The promise is deliberately small: ship from China, don't worry about the rest.

The approval gate is the product, not a safety blanket bolted on afterwards. The agents watch, read, decide and write the full message. Every send that leaves the company to a customer, a partner or a broker can be held for a human tap, and anything involving money, a restricted product or a commitment to a third party always is.

That is also why some quotes deliberately come back without a number. If the goods look like batteries, food or chemicals, automated pricing is withheld on purpose and a person reviews the shipment before anyone is quoted. An agency that will not quote a shipment it cannot legally move is more useful than one that quotes everything and apologises at customs.

The agents are Python on the Claude API, with cheap fast models doing routine classification and stronger models reserved for genuinely hard calls like an ambiguous HS code. Events run through AWS Lambda with SQS and SNS, state lives in Supabase, documents are generated as PDFs and stored in S3, and the customer-facing site is a Next.js app. Partner coordination with the China side runs over WeChat Work, broker communication over email, approvals over Telegram.

None of those are exotic choices. The interesting part is not the tools, it is wiring them into agents that can carry a shipment from quote to delivery with a single person supervising. That engineering problem, not the freight, is what I find worth writing about, and it is the same territory I work in as an AI consultant in Barcelona.

The company is operating. Quotes go out, shipments move, customs entries get filed, cartons land in Amazon FBA warehouses and at residential doors in the United States. Payment is by ACH or wire to a US bank account, with nothing due until the cartons are weighed at the China warehouse, so a customer is never paying against an estimate.

It was built manual-first on purpose. Take real shipments, feel the real friction, then automate each step once it is understood. That is the path Flexport's founder took, brokering by hand before writing code, and it is the reason the automation here is specified by awkward customs questions and warehouse miscommunications that actually happened. Guides written from those shipments are published for importers at plainfreight.com/guides.

Next is depth rather than breadth: a sharper HS classifier, fully automatic document generation, a customs pre-check that catches common filing errors before submission, and a landed-cost view that includes every fee. If you import from China, the fastest way to see where it stands is to ask it for a price.