# Parsing the Epstein Files for JMail — Palak Agarwal & Omar Alhait, Reducto

## Executive summary

The presentation details Reducto, an agentic document platform designed to convert messy, unstructured enterprise documents (such as scanned files, handwritten logs, and heavily redacted PDFs) into structured, usable data for AI applications. The core workflow involves a multi-stage pipeline: parsing, splitting, extracting, editing, and classifying. The platform uses a rasterization-based approach and an agentic OCR framework to handle complex layouts and reading order issues, exemplified by the JMail and JFlights case studies, which processed millions of pages of difficult government and personal records.

## Key takeaways

- Reducto's Core Functionality: Reducto acts as a pre-model layer, handling complex layouts and post-processing so that downstream models can focus on product logic. Its five core capabilities are: Parse, Split, Extract, Edit, and Classify. (0:00)
- Handling Unstructured Data: The platform addresses the challenge of unstructured documents (scans, rotations, charts) by using a rasterization-based approach, mimicking human visual understanding rather than relying solely on underlying embeddings. (0:00)
- The JMail/JFlights Case Study: The JMail product, which simulates logging into an inbox like Jeffrey Epstein's, required Reducto to parse over three million pages of redacted PDFs and messy handwriting. The JFlights product specifically handled complex flight logs, demonstrating the pipeline's ability to manage diverse document types and redactions. (2:02, 5:27)
- Schema Generation and Enhancement: Users can generate a JSON schema for extraction. While a quick schema can be generated via a simple LLM call, an enhanced schema generation is recommended, as it tailors the schema description directly based on the output of the parsed documents, improving accuracy for nuanced data extraction. (10:14)

## Technical details

- Agentic OCR Framework: The system uses an agentic OCR framework where the model does not commit to a single answer. Instead, it flags areas of uncertainty, corrects them, and proceeds, improving robustness over single-pass methods. (0:00)
- Document Processing Pipeline: The pipeline involves multiple stages: Parsing (the foundation), which reads the document structure; Splitting (dividing documents into regions); and Extraction (applying a schema to pull specific data). This sequence allows for complex, tailored data extraction. (10:14)
- Data Integrity and Redaction: The platform is designed to handle heavily redacted material. The speakers noted that simply drawing a black box over text does not guarantee full redaction, requiring careful processing to ensure data integrity and compliance. (12:30)
- Post-Processing and Entity Resolution: Advanced functionality includes post-processing steps, such as citation modeling, which links extracted data back to its source location within the document. (14:50)

## Practical implications

- For build engineers, Reducto provides a robust, pre-processing layer that can be integrated into data pipelines to ingest and structure highly variable, real-world document inputs, reducing the complexity of downstream ML model training.
- The ability to define and iterate on a JSON schema and test it against a live document corpus allows for rapid prototyping and validation of data extraction requirements.
- The architecture supports scaling from small, sample-based evaluations (vibe eval) to large, systematic benchmarks, which is crucial for productionizing data ingestion systems.

## Topics

Document AI, NLP, OCR, Schema Design, Data Pipelines, Agentic Systems, Reducto, JMail, JFlights, Reducto Studio

Source: https://www.youtube.com/watch?v=HGfsfKaVuGY