TaxHelper

Taxes are a nightmare for international students. Our AI assistant replaces rigid PDFs with a simple chat, using OCR and auto-generation to turn a stressful legal chore into a quick conversation.

Comment

Inspiration

Our project started from a frustrating personal experience. Earlier this semester, we were required to file federal taxes. This was our first tax season and we met series of difficulties, such as confusion by terminology and vague description on documents to return. So we spent hours navigating a maze of passport numbers, I-94 travel records, and Social Security details. One of our team members almost overpaid by hundreds of dollars due to a single minor mistake.

While our U.S. resident classmates could simply turn to their families for help, we realized that international students are often left to figure out these complex legal requirements entirely on their own. Even though our university provides great resources like Sprintax (an online tax compliance software), it’s an English system still with confusion to international students like us. What’s worse is that some of our friends at other schools told us they have almost no support at all. Some schools only send tax season email and then let students to do it on their own. This gap inspired us to develop a more intuitive tool—designed by international students, for international students—to make the filing process smooth, accurate, and accessible for everyone.

What it does

Our product is a free, open-access web platform designed to demystify the U.S. tax filing process for international students. It bridges the gap between complex IRS requirements and the specific needs of first-year F-1 visa holders.

1. Interactive AI Tax Consultant

Users interact with a specialized AI Tax Consultant through a clean chat interface. This assistant helps users:

  • Understand complex tax terminology.
  • Correctly determine their taxpayer status (specifically focusing on Non-Resident Alien identity).
  • Navigate the unique requirements of being a student filer in the U.S.

2. Streamlined Document Checklist

The platform provides a clear, personalized list of required materials. Users simply follow the checklist and upload their documents (such as W-2 forms) directly to the site.

3. Automated Form Generation

Once the data is collected, our backend takes over:

  • Auto-Classification: The system determines whether the user needs Form 8843 or Form 1040-NR.
  • Auto-Population: Our engine maps the extracted data directly into the official IRS templates.
  • Ready-to-File Output: Users receive a completed PDF that can be immediately printed and mailed to the IRS.

How we built it

1. Frontend & User Interface

The frontend serves as the primary data collection hub. It was built with two main focuses:

  • Dynamic Chat Consultant: We designed a structured interview flow that guides users through the specific questions required for IRS compliance.
  • Data Framework: Beyond the UI/UX, the frontend acts as a data pre-processor, capturing essential identity and residency information (specifically for Form 8843) and packaging it into a structured format for the backend. ### 2. OCR & Document Intelligence

To minimize manual entry, we integrated a sophisticated document processing layer:

  • W-2 Data Extraction: We utilized the MiniMax API to perform Optical Character Recognition (OCR) on W-2 forms.
  • JSON Transformation: The system identifies key financial fields and converts raw image/PDF data into a clean JSON format, ensuring seamless integration with our backend logic. ### 3. Backend & Logic Engine

The backend acts as the "brain" of the application, handling classification and document synthesis:

  • Tax Form Classification: The system evaluates user data passed from the frontend to determine filing requirements—specifically, whether the user needs Form 8843 (Statement for Exempt Individuals) or Form 1040 (U.S. Individual Income Tax Return).
  • Automated PDF Generation: Using predefined templates, the backend maps the JSON data into the official IRS PDF fields and returns a completed, ready-to-file document to the user. ### 4. Development Methodology (The AI Stack)

We utilized a multi-model Vibe Coding strategy to accelerate development. By providing detailed requirements to several Large Language Models (LLMs), we were able to generate robust structural frameworks quickly:

  • AI Models Used: Grok-4, Gemini Pro 2.5, and MiniMax 2.7.
  • Human-in-the-Loop Refinement: After the AI generated the core logic, we conducted extensive manual testing and "fine-tuning" on the logistics of the identification questions to ensure the tax logic remained airtight.

Challenges we ran into

The development process was defined by a series of iterative breakthroughs, where each technical hurdle forced us to refine our logic and system architecture.

1. Data Logic & Conversational Flow

Our primary challenge lay in the Taxpayer Identification Process.

  • Initial Data Gaps: In the early stages, our identification flow missed critical compliance information, such as specific visa types and residency history.
  • The 8843 Dilemma: We realized that the "Students Part" of Form 8843 requires specific manual entries. Implementing this created a significant debugging burden; we had to build robust error-handling to account for user typos and ensure the "logical branching" of the question set remained airtight through constant debugging. ### 2. OCR & Document Intelligence

Building the W-2 Tax Form Filler involved a frustrating cycle of troubleshooting.

  • The Invisible Error Loop: We spent significant time trapped in a generic "data extraction failed" loop. The breakthrough only occurred when we analyzed the server-side terminal logs, which revealed a persistent API error caused by an incorrect model name in our minimax.ts file.
  • Optimization: After overcoming syntax and compilation blockers, we faced low confidence scores (sometimes 45%) where the AI missed basic visible info. We had to pivot from just establishing a connection to actively optimizing extraction quality. ### 3. PDF Integration & Mapping

As the lead for backend integration and PDF generation, the primary hurdle was the "last mile" of data entry.

  • The "Blind" LLM Issue: A major technical roadblock was the program’s inability to accurately recognize specific coordinate locations within the PDF templates. The LLM essentially failed to "see" the PDF structure clearly enough to map the data.
  • The "Reference Point" Solution: To solve this, we developed a workaround where the AI first fills in all possible blanks as reference points. This provided the system with a "map" of the document, allowing it to accurately locate and fill in the correct answers in the proper fields.

Accomplishments that we're proud of

  • A Conversational Tax Solution: We have built a platform that replaces the traditional, headache-inducing PDF editing process with a streamlined Chatbox Interface. Instead of manually squinting at tax form fields and typing into rigid PDF boxes, users can now complete their entire filing through a natural conversation.
  • End-to-End Automation: We successfully integrated a complex pipeline—from OCR-driven W-2 data extraction to automated backend PDF generation. Seeing the system take raw user input and output a filled, ready-to-file tax form without any manual document editing was a major technical milestone for our team.
  • Lowering the Barrier for Non-Native Speakers: By combining LLM intelligence with a structured chat flow, we created a tool that doesn't just fill forms but guides users through them. We take pride in building a solution that makes U.S. tax compliance accessible and less intimidating for the international student community.

What we learned

The hackathon was more than just a technical challenge; it was a journey of shifting our perspective from "building a cool tool" to "solving a real human struggle."

Driven by Real-World Curiosity

Before the hackathon began, we went through a relentless cycle of brainstorming and discarding AI agent ideas. We realized that while AI is becoming ubiquitous, the specific tax struggles of non-native speakers remain a massive hurdle. This process taught us that true innovation starts with curiosity about a real problem; finding a gap in our own lives as international students gave us the self-driven motivation to build something that actually matters.

Engineering from the User’s Perspective

We gained a comprehensive understanding of the full-stack pipeline—from a raw concept to a functional website. Developing the frontend taught us that coding is as much about empathy as it is about logic. By repeatedly simulating the user journey, we realized that a "seamless" experience isn't an accident; it is the result of countless iterations, micro-adjustments, and debugging sessions focused on making the process feel intuitive for someone who is already stressed by taxes.

The Criticality of Technical Synchronization

One of our biggest takeaways was the importance of team collaboration over mere task-splitting. We learned that the "data handshake" between the frontend and backend is the heartbeat of an AI project. If the data captured from the user doesn't perfectly align with what the backend logic expects, the system fails. This required us to stay in constant sync to ensure that our individual parts—from OCR extraction to PDF generation—formed a single, functional loop that passed every test.

What's next

Our vision is to evolve from a text-based chat into a high-precision, interactive tax ecosystem that handles the complexity of international filings with zero friction.

  • Enhanced UI Interactive Plugins: To eliminate input errors and simplify the user experience, we plan to move beyond text-based Q&A by integrating specialized functional components:
    • Map & Location Integration: Much like a delivery app, we want to integrate map APIs so users can select addresses via search or pins, ensuring perfectly formatted data for the IRS.
    • Smart Calendar Components: For tracking "days of presence" and entry/exit dates, we will implement intuitive date pickers to ensure chronological accuracy.
    • Structured Selection Logic: For critical data like Visa Type, we will replace open-ended text with dynamic dropdowns and selection chips, ensuring the backend logic receives standardized, error-free input.
  • State-Specific Compliance Modules: U.S. tax laws vary significantly by state. We plan to build dedicated logic for states with high international student populations (such as California and New York), automatically generating the necessary state-level forms alongside federal filings.
  • Historical Record & Continuity Management: Tax filing is a multi-year commitment. We aim to implement a secure "Tax Vault" that stores past filings. This allows the AI to automatically track residency history and apply tax treaty benefits consistently year-over-year, making the second-year filing process as simple as a one-click confirmation.

Built With

Share this project:

Updates