Solution
Blue J built a tax-research engine on a Retrieval-Augmented Generation (RAG) system that pairs GPT-4.1 with a proprietary library of millions of curated documents, including authoritative primary sources and expert commentary such as Tax Notes. When a user asks a question, the platform retrieves the relevant statutes, regulations, rulings, case law, and commentary, then uses GPT-4.1 to synthesize a cited answer designed to read like guidance from a trusted colleague. Blue J wrapped that flow in a feedback and evaluation loop: every answer can be disputed through a disagree button, GPT-4.1 clusters issue types and likely root causes across thousands of feedback points, and a benchmark suite of 350+ prompts across U.S., Canadian, and U.K. tax law gates model updates. This closed-loop architecture let the company launch quickly, update answers within hours of major legislation, and scale trusted tax research across multiple countries.
Data flow
A tax professional submits a question, the RAG system retrieves the matching primary sources and commentary from Blue J's curated library, and GPT-4.1 synthesizes a cited answer. If the user flags an issue through the disagree button, GPT-4.1 categorizes the feedback, clusters related problems, and helps the product and tax research teams tune retrieval and prompts before updated answers are pushed back to production.
Solution architecture
2 components · 2 layers - Compute
- GPT-4.1 Synthesizes cited answers, follows domain-specific instructions, and triages user feedback to improve the product.
- Orchestration
- Retrieval-Augmented Generation (RAG) system Retrieves the relevant tax authorities and commentary from Blue J's curated document library before answer generation.
- Every answer includes a disagree button for structured feedback capture
- GPT-4.1 analyzes feedback points by issue type, tax topic, and likely root cause
- Model releases are tested against 350+ prompts spanning U.S., Canadian, and U.K. tax law
- RAG architecture combines GPT-4.1 with a proprietary library of millions of curated documents
- Users saw updated answers within hours of a major 2025 U.S. tax bill being signed
"Blue J's tax research solution is built using a Retrieval-Augmented Generation (RAG) system, combining GPT-4.1 with a proprietary library of millions of curated documents, including authoritative primary sources and expert commentary from sources like Tax Notes."
"GPT-4.1 powers this triage layer, analyzing thousands of feedback points, clustering related issues, and helping Blue J's product and tax research teams focus their efforts where they'll have the biggest impact."
"More than 70% of users log in weekly, saving 2.7 hours per user per week on research and client communication."
"Today, more than 70% of users log in weekly, with a disagree rate of fewer than 1 in 700 responses."
"Within hours of the bill being signed, users were seeing updated answers in production."
