jc2744-fintech/taxguard-xai

Interpretable AI framework for compliance risk screening on IRS Form 990 filings — transformers, GNNs, and temporal models with built-in explainability.

What it solves

TaxGuard XAI is a research-grade framework designed to help auditors and researchers prioritize the review of public IRS Form 990 filings. Instead of reviewing every filing manually, it uses machine learning to identify records that may deserve human verification based on financial anomalies, narrative patterns, and organizational relationships.

How it works

The system processes IRS TEOS XML data or ProPublica API data through a multi-layered pipeline:

  1. Data Ingestion: Normalizes raw filings, relationship graphs, and officer data.
  2. Feature Engineering: Extracts structured financial ratios, temporal trends (year-over-year changes), and narrative text.
  3. Modeling Modes:
    • Anomaly Discovery: Uses Isolation Forest to find unusual records relative to a reference dataset.
    • Weak Supervision: Applies 11 auditable labeling functions to generate hypotheses for review.
    • Supervised Research: Trains models on human-reviewed labels using backtesting and out-of-time metrics.
  4. Explainability (XAI): Every score is accompanied by evidence, such as token attribution for text, edge attention for graphs, and integrated gradients for temporal data.
  5. Fusion: Combines signals from text, graph, and structured data using reliability-aware fusion.

Who it’s for

It is built for tax researchers, auditors, and data scientists focusing on non-profit governance and financial transparency.

Highlights

  • Real-world Data Integration: Includes official IRS bulk download paths and ProPublica API support.
  • Audit-First Design: Maintains a strict separation between model signals and final human-verified findings.
  • Multimodal Analysis: Optionally supports Transformers for text, GATv2 for relationship graphs, and Temporal Convolutional Networks (TCN) for longitudinal data.
  • Transparent Outputs: Generates HTML review bundles containing the specific reasons, ratios, and evidence behind every risk score.

Related

  • Project
    rollingSirius/equity-research-skill

    An AI-powered equity research framework that transforms AI agents into institutional-grade analysts by enforcing rigorous valuation, forensic accounting, and industry-specific research protocols.

  • Dispatch
    OpenAI Privacy Filter: Building Scalable PII Detection Web Apps

    OpenAI has released Privacy Filter, an open-source 1.5B-parameter PII detector capable of labeling eight categories of sensitive data across a 128k context window.

  • Project
    Trusted-AI/AIX360

    An open-source library providing a comprehensive set of algorithms and metrics to improve the interpretability and explainability of machine learning models and datasets across various data modalities.

  • Project
    katanaml/sparrow

    An API-first platform for enterprise document intelligence that converts invoices, receipts, and statements into structured JSON using Vision LLMs and agentic workflows.

  • Dispatch
    xAI Grok Collections API Release

    xAI has launched the Collections API, a managed knowledge base service that enables developers to build RAG applications by uploading datasets for precise semantic, keyword, and hybrid search.