AI Customs Clearance

Inspiration

The inspiration for this project's solution was the lack of automation in terms of supply chains. Before a product is shipped it has to go through customs, before shipping someone has to check paperwork. Which are the invoice ( what was bought ), the packing list ( what is actually in the boxes ) and the billing ( the shipping receipt ). This process is slow, monotonous and must be streamlined for productivity. One mistake for example a typo in quality could cause delay in shipments. AI could be implemented to read, spotting and responding to these mistakes. Our question in implementation is “ would you let AI release a shipment through customs on its own? “. No, most people’s answer is correct. The agent does most of the work with strict guidelines, it suggests where as the people decide.

What it does

Our AI agent reads a shipment and its respective documents and looks for possible errors Thinks of the problem and proposes a plan to resume productivity and shipments. Asking such answers like: "ask the warehouse manager to confirm the real quantity". Checks the plan with the rulebook: assets a safe plan and actions to go ahead . Risky ones require human approval or blocked outright. While any plan requires human improval Waits dor human or correct people for approval: once approval it carries out the plan and finishes the job. Records every action in a record that can’t secretly be edited.

How we built it

How the system works together: The brain: The smart assistant who reads documents and writes suggestions. Which is really is: a language model supported by openAI GPT and IBM Bob AI

The Rules: a strict manager that reviews and signs off decisions. What it really is: plain computer code, no AI so it can’t be talked out of a decision.

The Memory: a notebook: where every page is sealed to the page before it. It remembers previous shipments. What it really is: a database plus a “harsh chain”. Or the Agent follows a fixed map of steps where each steps has to pass the rulebook first. It follows the logic of reason, plan, policy check, act, verify, and cheer.

The Window: A dashboard in the front end that houses the user experience. WHat it really is showing everything live.

Human decision is crucial to what we built. Human decision: allows the agent to go ahead, a person must approve. A person may block the agent’s verdict.

Challenges we ran into or faced was:

Each of these taught us something:

Approvals quietly failed. At first, approval requests never actually reached the system because of a database error, and no one noticed. The agent looked like it was "waiting for a human", but there was nothing for a human to approve. We fixed it, and our tests now run against a database that behaves like the real one. The agent forgot what it was doing. Each time it woke up, it started from zero. We gave it a memory, so after people approve, it picks up exactly where it stopped. The AI was silently replaced by a backup. Our AI key stopped working because it had been accidentally uploaded to a public code site and was then switched off. The system quietly used pre-written backup answers instead, so the screen looked normal but no real AI was involved. We now show a clear label on screen: "Live AI" or "Offline fallback". A loophole in the rulebook. We found that someone could send a request saying "risk score = 0" and skip the rules entirely. We closed it: risk is always calculated by the system and never accepted from outside. Hidden bugs in the risk math. Two subtle bugs were making some dangerous actions look safer than they were. The fix was to score each action on its own. "It works on my computer." When we first published the website, it worked only on our own laptop. A leftover development setting made every visitor's browser look for data on their own computer. Other people saw only an error. Some networks blocked our web address. We added a second, plain-IP address as a backup. Judges found the screen confusing. The ideas were strong, but a first-time visitor didn't know where to click. That feedback led to the tour, the help tips, the user guide and the reset button.

Accomplishments that we're proud of

We are proud of building a human first and in charge AI agent that could streamline custom checks for world trade. Using IBM Bob open AI to build this agent from scratch creating a usable dashboard that anyone can use and having the AI follow guidelines for trade Also getting the agent to work for every browser or local channels so it could be used for everyone.

What we learned

AI proposes, rules decide." You can trust AI with a lot of work, as long as the final say over dangerous actions belongs to simple, predictable rules and accountable people. Assume the AI can be fooled. Instead of hoping the AI resists tricks, we designed the system so that being tricked does no harm. Things that fail silently are the most dangerous. Several of our worst bugs looked fine on the surface. Showing honestly what is really happening, such as "Live AI" versus "fallback", builds more trust than a screen that always looks fine. Test like a stranger. Our own tests passed while real visitors couldn't use the site. Now we test from a fresh browser, the way a new user would. Keep secrets secret. Passwords and keys must never be uploaded with the code. Clear explanations are part of the product. A good idea only counts if people can understand it in a minute.

What's next for AI Customs Clearance

Use IBM watsonx Granite alongside a second AI model, and send a case to a human whenever the two disagree. Role-based logins, so each person can approve only the decisions that belong to them. Give each visitor their own private copy of the demo, so several judges can try it at the same time without getting in each other's way.