NEPA Clean Energy Review Analysis

Built a Python BERT+LLM pipeline for the Clean Air Task Force to extract and classify 120,000+ federal environmental review documents, enabling evidence-based permitting reform recommendations.

NEPA Clean Energy Environmental Review Analysis

Client: Clean Air Task Force  |  Jan 2026–Present

The Research Problem

Federal clean energy projects — wind, solar, transmission — must complete environmental reviews under the National Environmental Policy Act (NEPA) before breaking ground. These reviews can take years and run to thousands of pages, creating a significant bottleneck for the clean energy transition. But because NEPA documents are unstructured text scattered across federal agencies, no one had systematically measured how long reviews actually take, which agencies are the slowest, or which project types face the steepest delays.

The Clean Air Task Force — a leading clean energy policy nonprofit — needed that evidence base to make credible, agency-specific permitting reform recommendations to policymakers and funders.

Client Work: From Research Questions to Deliverables

Working directly with CATF researchers, I led the full arc of the project:

  • Scoped the research questions collaboratively with CATF, translating high-level policy goals (“what’s slowing clean energy permitting?”) into measurable, answerable questions about review timelines, agency behavior, and project characteristics
  • Identified and acquired the data, integrating the PNNL NEPATEC 2.0 corpus (120,000+ NEPA documents) with Federal Register API data and targeted web scraping to build a comprehensive project-level dataset
  • Delivered phased findings across six structured deliverables, each mapped to a specific CATF research priority — from baseline timeline benchmarks to multi-agency coordination patterns
  • Communicated results to non-technical stakeholders, presenting findings and strategic recommendations in stakeholder presentations and a public-facing analytics site
  • Built reusable infrastructure so CATF can extend the analysis independently as new documents are added

NLP & Machine Learning Pipeline

The core technical challenge was turning raw, inconsistently formatted text documents into structured data at scale.

  • Data ingestion — loaded the PNNL NEPATEC 2.0 corpus and supplemented it with Federal Register API data and targeted web scraping, preprocessing everything into per-source parquet files for efficient querying
  • Project classification — applied regex and ML classifiers to categorize 20,000+ projects by technology type (solar, wind, transmission, storage), review process (CE, EA, EIS), and installed capacity
  • BERT extraction — trained BERT classifiers to pull review start/end dates, categorize review type, and flag timeline milestones from free-form document text
  • LLM adjudication — for extractions where BERT confidence was low, a second-pass LLM layer resolved ambiguity by combining model outputs with rule-based post-processing
  • Analysis and reporting — structured parquet outputs feed into per-deliverable R scripts that produce figures and tables; Quarto renders everything into HTML reports published to the project website

The full pipeline is reproducible end-to-end and built on DuckDB (Phase 2) for scalable, query-efficient processing of the full corpus.

Left: Clean energy NEPA review filings by year, extracted from 120,000+ documents. Right: Distribution of projects by technology type, classified using ML and regex pipelines.

Public Deliverables

Findings are published in two public-facing tools designed to give policymakers, funders, and advocates direct access to the analyses and underlying documents:

  • Project Website — interactive HTML reports with figures, tables, and narrative summaries of each deliverable
  • HuggingFace Document Browser — a Streamlit app for exploring individual NEPA documents, source classifications, and extracted metadata
Agency coordination patterns in multi-agency NEPA reviews, visualized as a Sankey diagram in the public analytics site.

Skills & Methods

Python  BERT  LLMs  NLP  DuckDB  R  Quarto  Streamlit  Federal Register API  Web scraping  Regex  Machine learning  Data visualization