Tariffind — See the Invisible Tax
What inspired us
The inspiration for Tariffind came from the chaotic and rapidly changing landscape of US trade policy. During the Tech@NYU Startup Week Buildathon (Feb 20-22, 2026), a unique set of headlines broke: The Supreme Court struck down tariffs in the morning, and the administration signed new ones by the afternoon. For the average consumer, this geopolitical whiplash translates to one thing: an invisible tax on everyday goods. We realized that shoppers have zero visibility into how much of a product's price is artificially inflated by layered trade policies. We wanted to build a tool that makes the invisible tax visible, empowering consumers to understand what they are actually paying for and helping them find cheaper, less-taxed alternatives.
How we built our project
We built Tariffind using a modern, AI-first architecture to handle the extreme complexity of international trade codes and apply cascading tariff logic on the fly.
Our stack integrates several powerful tools to make this happen:
- Claude Agent SDK: Acts as our core intelligence layer. When a user inputs a product name or URL, the Claude-powered Product Classifier instantly translates the informal product description into the correct Harmonized Tariff Schedule (HTS) code and identifies the likely country of origin.
- FastAPI & Python: The backend is driven by a robust Tariff Engine that calculates the compounding effects of multiple active trade policies.
- Lovable & React: We used Lovable to generate a clean, reactive frontend that visualizes the price breakdown instantly for the user.
Challenges we faced
Navigating the labyrinth of the US Harmonized Tariff Schedule is notoriously difficult. Our biggest challenge was data structuring and real-time calculation. The HTS contains tens of thousands of deeply nested categories, and layering the different geopolitical tariffs (especially conditional ones like Section 301 for China vs. blanket ones like post-SCOTUS Section 122) required building a strict precedence and applicability engine.
Additionally, mapping natural language (e.g., "Samsung 65 inch QLED TV") to a highly specific legal HTS code (like 8528.72.64) is mathematically inexact. We had to carefully prompt and tune the Claude Agent SDK to reliably output structured JSON with the correct classification without hallucinating trade codes. We also had to build an active keyword fallback system just in case the AI failed or timed out during the lookup.
What we learned
Building Tariffind taught us that trade policy isn't just an abstract macroeconomic concept—it's a direct, measurable tax on consumers. We learned an immense amount about how global supply chains are categorized, tracked, and taxed by customs agencies.
From a technical perspective, we learned how to effectively compose different AI and large-scale data tools. Bridging the gap between Databricks for heavy data pipelines, Claude for nuanced language reasoning, and FastAPI/React for real-time user feedback showed us the true power of full-stack AI orchestration. The event also pushed us to rapidly integrate new tools like Lovable for frontend generation and ElevenLabs for voice accessibility, demonstrating how quickly an infrastructure can be spun up when leveraging modern platforms.
Log in or sign up for Devpost to join the conversation.