AI-Powered Simplification of the U.S. Tax Code
🧠 Project: AI-Powered Simplification of the U.S. Tax Code
Repository: github.com/candrasick/ai_tax_agent
Overview
In this capstone project, I built a suite of AI agents using Google’s Gemini 1.5 Pro to tackle one of the most complex regulatory systems in existence: the U.S. tax code. The goal was ambitious — to simplify the legal text, maintain revenue neutrality, and surface low-impact, high-complexity candidates for deletion or rewrite.
The first full pass through 2,100+ sections resulted in a 60% reduction in length, compressing over 7,000 pages to fewer than 3,000. And this is just the beginning.
🎬 Video Summary
You can find a video summarizing the project here:
🔍 Key Features and Capabilities
🧾 Exemption Extraction
Gemini was used to parse each section of the U.S. Code and extract tax exemptions — often deeply embedded or cross-referenced — as discrete entities. This enabled clearer mapping of tax advantages and loopholes to statutory sections.
🧪 Integration Testing of Heuristics
Using Gemini as a cognitive test harness, I verified custom logic against form metadata and code references. This allowed for automated regression checks during agent design and helped tune the heuristics used for section-to-stat linkage.
📊 Parsing Unstructured PDF Statistics
Gemini’s multimodal capabilities were used to convert IRS statistics — often locked in PDF tables — into structured JSON. This enabled dynamic linkage between form line items and statutory authority, empowering more accurate impact estimation.
📉 Complexity Scoring
Each section was scored based on:
- Legal language complexity (measured directly)
- Z-scores for section length
- Amendment count
- IRS bulletin references
Gemini synthesized these into a final complexity_score, which guided subsequent rewrite decisions.
💰 Impact Cataloging
The agent estimated financial and entity impact using:
- Tax statistics from the IRS
- Line item instructions
- Linked form references
This allowed the agent to identify high-impact areas of the code and prioritize retention or careful editing.
🧭 LLM-Grounded Decision Making
The agent made edit decisions (simplify, redraft, delete, keep) by reasoning across both complexity and impact. Sections with low complexity and low impact were often deleted. High-impact sections were simplified or redrafted conservatively.
🧠 ChromaDB Vector Search
ChromaDB was used to semantically retrieve related sections, exemptions, and form metadata. This gave the LLM critical context and allowed deeper insight into how a section connects across the tax ecosystem.
💡 Results
- Reduced text length by ~60% in the first pass.
- Preserved or improved readability across over 1,000 sections.
- Maintained revenue neutrality within acceptable bounds.
- Established an agentic feedback loop that can continue refining the code across future iterations.
💾 Open Source
All code, prompts, and data are open-sourced and available at: