Can an AI agent tell a trustworthy supplier from a convincing fake?
SupplyCorridor is an early-stage concept for more transparent international sourcing. The associated research project proposes a controlled evaluation of how AI agents respond to misleading supplier information and whether safeguards actually work.
International trade runs on information. So do AI agents.
Buyers often need to compare websites, supplier claims, registration documents and third-party records across borders. Changing trade rules and logistics requirements add complexity. If an autonomous agent treats unreliable information as fact, it can recommend the wrong counterparty or act before risks have been checked.
The research question extends beyond international trade: how should an agent handle evidence it cannot confidently verify?
SupplyCorridor
The broader platform concept explores sourcing relationships between buyers in the UK and EU and manufacturers in India, with supplier discovery, structured requests and more trustworthy information exchange.
Independent evaluation
The proposed research prototype would test information integrity, decision logging, uncertainty thresholds and human review in a controlled environment using synthetic or benign public test data.
Test the weak points. Measure the difference.
Use synthetic supplier pages with inconsistent registrations or invented credentials.
Present competing records about business details, location or compliance.
Test whether page text causes an agent to disregard checks or hide uncertainty.
Measure decisions made before adequate evidence or human approval is available.
What's been done.
What's still ahead.
Transparency is essential when asking others to trust a new technical project. The initial business work is distinct from the proposed AI safety build.
Completed or explored
An initial SupplyCorridor website was built and previously published using Replit, according to the founder; its deployment is currently inactive. Early work has examined international sourcing, logistics, trade agreements and supplier information quality. The founder has also completed relevant postgraduate research.
To be developed and tested
A dedicated agent evaluation environment, synthetic adversarial cases, a safeguard comparison, independent technical review, documented metrics and a public research report. No functioning AI safety benchmark or live pilot is claimed at this stage.
Focused first milestone
Define threat model, scoring rules, synthetic cases and technical architecture.
Implement controlled agent workflow, verification checks and review triggers.
Measure baseline failures versus safeguard performance; document limitations.
Seek technical feedback and prepare a reproducible report and oversight checklist.
Grounded in international business and emerging technology.
Project led by Vaibhav Dureja, founder of Mooncommerce Limited. MSc International Business, Birkbeck, University of London (Merit), with dissertation research on AI, blockchain and IoT.
A 2026 SSRN working paper on emerging technologies and UK business transformation. Publicly shared research, not a peer-reviewed publication.
Early-stage product scoping at Mooncommerce, including digital marketplaces, cross-border sourcing and information verification.
Controlled test data rather than private supplier records; publication of methods and outcomes where licensing and safety permit.
Interested in practical AI safety evaluations?
I'm exploring technical collaboration, independent review and connections with researchers working on autonomous-agent reliability and oversight.