AI Engineer · Published 2026-10-11

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

privateReview: unreviewedConfidence: high
reductoocrparsingextractiondocumentsdocument-parsingrasterizationschema-generationagentic-ocringestion-pipelinemultimodal
Open on YouTube ↗

Summary

Overview

  • Speaker: Palak Agarwal & Omar Alhait
  • Channel: AI Engineer
  • Main topic: Document ingestion and parsing pipeline for messy enterprise data and unstructured documents using Reducto
  • Purpose: Explain how Reducto solves complex document ingestion challenges for AI teams, demonstrate real-time parsing of messy historical documents (Epstein files), and discuss the technical trade-offs of schema generation, OCR post-processing, and document layout handling. Palak Agarwal and Omar Alhait from Reducto present how their document ingestion and parsing platform handles complex, unstructured, and messy enterprise documents—specifically showcasing how they parsed over 3 million pages of Epstein flight logs, scanned forms, handwritten notes, and government files to power JMail in just five hours. They detail their rasterization-based approach, agentic OCR frameworks, and schema-driven extraction pipelines.

Topic Map

Introduction to Reducto and Document Ingestion

  • Explanation: Overview of Reducto as an agentic document platform helping teams turn messy unstructured documents into AI-ready data.
  • Key claims:
    • Reducto helps teams turn messy unstructured documents into data AI can use.
    • Over $108M raised in funding, processing over 2 billion documents across major customers.
  • Examples:
    • Processing documents for Harvey, Scale, and top global hedge funds.
  • Terminology:
    • rasterization-based approach
    • agentic OCR
    • post-processing
  • Why it matters: Enterprise documents are notoriously messy (scanned, rotated, handwritten, charts), requiring a dedicated pre-model layer before LLMs can ingest them effectively.

The JMail Case Study & Parsing Flight Manifests

  • Explanation: Demonstration of parsing complex handwritten and scanned passenger manifests from the Epstein files using Reducto's studio interface.
  • Key claims:
    • JMail was cloned from Gmail in 5 hours following the release of Epstein files.
    • Reducto parsed 3M+ pages of scanned PDFs, handwritten notes, and government files behind JMail.
  • Examples:
    • Extracting passenger manifests with handwritten annotations, flight numbers, and dates.
  • Terminology:
    • JMail
    • passenger manifest
    • extraction schema
  • Why it matters: Shows how real-world messy historical documents with complex layouts and handwriting can be automatically structured into clean JSON/schemas using a robust parsing pipeline.

Schema Generation and Extraction Pipelines

  • Explanation: Explaining the difference between standard extraction schemas and enhanced schema generation, and how parsing feeds downstream data products.
  • Key claims:
    • Extraction schema acts as the paintbrush on the Reducto canvas.
    • Reducto offers both standard schema generation and enhanced schema generation tailored to parsed document contents.
  • Examples:
    • Creating a resume parsing pipeline versus a flight history extraction pipeline.
  • Terminology:
    • schema generation
    • extraction pipeline
    • parse
    • split
    • extract
    • edit
    • classify
  • Why it matters: Proper schema design and document parsing allow developers to rapidly build specific downstream products without writing brittle custom regex parsers.

Key Points

Rasterization-Based Approach to Document Ingestion

  • Explanation: Reducto views documents visually like a human would rather than relying strictly on underlying embeddings.
  • Evidence: Handling rotated documents, charts, scans, and handwritten text directly through visual models.
  • Practical implication: Significantly improves accuracy on complex enterprise layouts compared to naive text extraction.

Agentic OCR Framework

  • Explanation: Rather than making a single pass and committing to an answer, the model flags uncertainty, corrects itself, and proceeds step-by-step.
  • Evidence: Framework logic built into Reducto's ingestion engine.
  • Practical implication: Reduces hallucinations and structural errors in extracted data.

Core Platform Capabilities

  • Explanation: Five core capabilities: Parse, Split, Extract, Edit, and Classify.
  • Evidence: Demonstrated live in the Reducto studio interface during the JMail walk-through.
  • Practical implication: Provides a comprehensive toolset for developers building document-heavy AI workflows.

Frameworks, Models & Processes

Reducto Document Processing Pipeline

  • How it works: First parses documents visually using rasterization and agentic OCR, then applies configuration schemas to extract structured data, followed by downstream product integration.
  • Components:
    • Parse
    • Split
    • Extract
    • Edit
    • Classify
  • When to use: When ingesting messy, unstructured, multi-format enterprise documents for LLM applications.

Examples & Case Studies

Parsing handwritten Epstein flight logs and passenger manifests in JMail.

  • Illustrates: Handling messy handwriting, complex tables, and degraded scans.
  • Lesson: Visual rasterization and agentic OCR successfully bridge the gap between messy historical paper records and structured AI data.

Actionable Takeaways

  • Immediate:
    • Sign up at studio.reducto.ai to test document extraction.
    • Create an API key and plug get-started prompts into your codebase.
  • Strategic:
    • Enterprise document ingestion requires dedicated pre-processing layers (layout analysis, OCR correction, schema mapping) before passing data to LLMs.
    • Visual document understanding outperforms raw text extraction on complex forms and tables.
  • Questions to investigate:
    • How does Reducto handle extremely large documents with thousands of pages at scale?
    • What are the latency benchmarks for hybrid OCR model pipelines?

Claims Worth Verifying

  • Reducto processed over 2 billion documents across enterprise customers. (metric)
  • JMail processed over 3 million pages of scanned PDFs and government files. (metric)

Notable Quotes

"We're an agentic document platform which basically means we help teams turn messy unstructured documents into data that AI can actually use accurately." (at 0:23) "JMail was originally built by Riley Walls and Luke Igel in five hours last November. That's when the Epstein files were released." (at 3:36) "your extract schema is sort of like the paintbrush on your reducto canvas" (at 16:58)

Compressed Summary

  • Reducto provides an agentic document platform for parsing messy unstructured files.
  • Demonstrated ingestion of 3M+ pages of Epstein files for JMail using visual rasterization.
  • Platform core capabilities include parse, split, extract, edit, and classify.
  • Offers both quick schema generation and enhanced schema generation for nuanced extractions.
  • Keywords: reducto, ocr, parsing, extraction, documents
  • Core insight: Reducto uses a rasterization-based visual approach and agentic OCR to convert messy, unstructured enterprise documents into clean, structured data for AI applications.

Core insights

5
Architecturemedium noveltymoderate evidence

Reducto treats document ingestion as a visual understanding problem — documents are viewed like a human would (rasterized), rather than relying strictly on the underlying text layer or embeddings. This reframes parsing as image/layout reasoning rather than text extraction.

Why it matters

It moves the hardest part of document ingestion (rotated pages, charts, handwriting, degraded scans) into a visual model boundary instead of a brittle text-extraction boundary, which is where most naive pipelines fail.

Generalization

For layout-heavy or handwritten inputs, grounding extraction in rendered visual representations can outperform pipelines that depend on the document's own text layer.

Reducto views documents visually like a human would rather than relying strictly on underlying embeddings.
Open source video
Handling rotated documents, charts, scans, and handwritten text directly through visual models.
Open source video
Mechanismmedium noveltymoderate evidence

The OCR layer is agentic: instead of a single pass that commits to an answer, the model flags uncertainty, self-corrects, and proceeds step-by-step within the ingestion engine.

Why it matters

Uncertainty-flagging plus iterative self-correction is a concrete mechanism for reducing hallucinations and structural errors, and it implies OCR should be a stateful loop, not a stateless transcription call.

Generalization

Extraction stages in agentic pipelines can be made more reliable by explicitly surfacing uncertainty and allowing a correction pass rather than emitting a final answer in one shot.

Rather than making a single pass and committing to an answer, the model flags uncertainty, corrects itself, and proceeds step-by-step.
Open source video
Architecturemedium noveltymoderate evidence

Document processing is decomposed into five orthogonal primitives — Parse, Split, Extract, Edit, Classify — exposed as a composable toolset rather than a single monolithic parser.

Why it matters

Separating layout parsing from splitting, extraction, editing, and classification lets developers compose document workflows and swap stages independently, instead of writing custom brittle parsers per use case.

Generalization

Complex ingestion problems are easier to reason about when split into discrete, composable stages with clear responsibilities.

Five core capabilities: Parse, Split, Extract, Edit, and Classify.
Open source video
Provides a comprehensive toolset for developers building document-heavy AI workflows.
Open source video
Mental Modelmedium noveltymoderate evidence

Ingestion is positioned as a dedicated pre-model layer (layout analysis, OCR correction, schema mapping) that runs before data reaches the LLM, rather than being folded into the LLM prompt.

Why it matters

It assigns responsibility for document cleanup to a separate preprocessing layer, keeping the LLM downstream of cleaned, structured data instead of asking it to compensate for messy inputs.

Generalization

Reliability of downstream reasoning improves when messy-format normalization is handled by a purpose-built layer with its own correctness concerns, not by the generative model.

Enterprise document ingestion requires dedicated pre-processing layers (layout analysis, OCR correction, schema mapping) before passing data to LLMs.
Open source video
Practicemedium noveltymoderate evidence

Schema generation can be tailored to the actual parsed document contents (enhanced) rather than fixed up front (standard), making the extraction schema a configurable artifact applied after parsing.

Why it matters

Choosing between a quick generic schema and a content-tailored one is a design lever that affects nuance capture versus setup speed; treating the schema as a post-parse artifact decouples parsing from the target data shape.

Generalization

Extraction quality can be improved by generating or adapting the output schema after inspecting the source content rather than committing to a static schema before parsing.

Reducto offers both standard schema generation and enhanced schema generation tailored to parsed document contents.
Open source video
your extract schema is sort of like the paintbrush on your reducto canvas
Open source video

Deep dives

5

Visual rasterization vs. text-layer extraction for messy documents

Research question

On which document classes (rotated scans, handwritten forms, charts) does a rasterized visual parsing pipeline beat text-layer extraction, and what is the accuracy gain per unit of added compute and latency?

Why

Determines whether the ingestion boundary should be an image/layout model or a string extraction step — the choice cascades into rendering infrastructure, cost, and per-page latency budgets.

They detail their rasterization-based approach, agentic OCR frameworks, and schema-driven extraction pipelines.
Open source video
Enterprise documents are notoriously messy (scanned, rotated, handwritten, charts), requiring a dedicated pre-model layer before LLMs can ingest them effectively.
Open source video
Source video

Uncertainty signaling and stopping criteria in agentic OCR

Research question

What uncertainty signal and stopping condition should an agentic OCR loop use to decide when to re-read or self-correct a span, and does the correction pass reduce structural error more than it costs in latency?

Why

Without a defined uncertainty signal and stopping rule, an agentic OCR loop is either useless (never corrects) or unbounded (corrects forever); both outcomes are unmeasurable in production.

They detail their rasterization-based approach, agentic OCR frameworks, and schema-driven extraction pipelines.
Open source video
Extracting passenger manifests with handwritten annotations, flight numbers, and dates.
Open source video
Source video

Post-parse schema generation (standard vs. enhanced) and schema validation

Research question

Does generating the extraction schema after inspecting parsed content measurably improve field-level accuracy over a pre-committed standard schema, and how can a generated schema be validated before it is trusted downstream?

Why

If the schema is a post-parse artifact, teams need a schema-critique step; otherwise enhanced generation just moves the brittleness from regex parsers to unvalidated schemas.

Reducto offers both standard schema generation and enhanced schema generation tailored to parsed document contents.
Open source video
Proper schema design and document parsing allow developers to rapidly build specific downstream products without writing brittle custom regex parsers.
Open source video
Source video

Scaling and cost of visual-OCR pipelines on thousand-page corpora

Research question

Where do per-page cost, latency, or memory become binding constraints when rasterizing and visually parsing multi-thousand-page documents, and what chunking strategy preserves cross-page context?

Why

The headline 3M+ page parse says nothing about the marginal cost of page 5000 in a single document; without that curve, capacity planning is guesswork.

Reducto parsed 3M+ pages of scanned PDFs, handwritten notes, and government files behind JMail.
Open source video
Over $108M raised in funding, processing over 2 billion documents across major customers.
Open source video
Source video

Composable ingestion primitives as a substitute for custom parsers

Research question

Does decomposing ingestion into Parse / Split / Extract / Edit / Classify reduce per-workflow custom code and error rates versus a monolithic per-document-type parser?

Why

The claim that teams can recombine primitives instead of rebuilding parsers is an engineering-economics claim that is directly falsifiable on code volume and defect counts.

schema generation, extraction pipeline, parse, split, extract, edit, classify
Open source video
Creating a resume parsing pipeline versus a flight history extraction pipeline.
Open source video
Source video

Article ideas

4

Your PDF Text Layer Is Lying to You: Move Document Ingestion Into the Visual Model

For messy enterprise inputs, rasterizing pages and reasoning over rendered layout beats trusting the embedded text layer, so the ingestion layer — not the LLM prompt — should own rotation, handwriting, and charts.

Angle

Before/after pipeline walkthrough showing where text-layer extraction silently loses information and where a visual render preserves it.

Source video

OCR Should Be a Loop, Not a Call

Treating OCR as a stateless transcription call bakes errors into downstream data; a stateful loop that flags uncertainty and re-reads suspect spans spends compute at the cheapest point in the pipeline rather than paying for LLM repair later.

Angle

Contrast single-pass transcription with an uncertainty-flagged correction loop, and argue about where the extra latency budget is best spent.

Source video

Generate the Schema After You Read the Document

Pre-committing an extraction schema trades nuance for speed; deriving or enhancing the schema from parsed content — with a validation step against gold fields — is the higher-fidelity default for heterogeneous corpora.

Angle

Design-lever argument: when standard schema speed is worth it versus when enhanced schema nuance pays off, plus how to grade a generated schema.

Source video

The Pre-Model Layer: Keep Messy Formats Out of the Prompt

Asking an LLM to compensate for scanned, rotated, or handwritten inputs wastes tokens and reliability; a normalization layer with its own correctness metrics keeps the model downstream of clean structured data.

Angle

Architectural responsibility argument: separating normalization from reasoning, and why 'just prompt harder' fails on degraded inputs.

Source video

Project ideas

4

Raster-vs-Text A/B Harness

beyond-evals

On rotated, handwritten, and scanned pages, a rasterized visual parsing pipeline yields higher field-level extraction F1 than a text-layer extraction pipeline at equal or lower cost per corrected field.

Proof of concept

Take a public corpus of scanned forms or FOIA PDFs, hand-label ground-truth fields, then run both a render-then-parse pipeline and a text-layer extraction pipeline over the same pages and score them with identical field definitions.

Measurement

Field-level exact match and F1 per document class, cost per corrected field, p95 latency per page, and the document classes where the accuracy gap flips direction.

Source video

Agentic OCR Correction Loop

movement-lab

Adding an explicit confidence signal and one bounded correction pass over low-confidence spans reduces structural and field errors on handwritten pages versus single-pass OCR, at a bounded latency increase.

Proof of concept

Wrap a vision-OCR stage with a self-reported confidence field, re-read only spans below a threshold once, then compare outputs against the same pages processed single-pass.

Measurement

Structural error rate (malformed output / field misalignment), field F1 delta, and added latency per page; sweep the confidence threshold to find the cost/benefit knee.

Source video

Post-Parse Schema Generator with a Schema Critic

new

Generating the extraction schema after parsing improves field F1 on heterogeneous corpora versus a fixed standard schema, and a schema critic can score proposed fields against gold with measurable precision.

Proof of concept

Parse a mixed sample corpus, auto-propose a schema, extract with it, and compare against a fixed-schema baseline; add a critic that matches proposed field names and types to a gold schema.

Measurement

Field F1 delta versus fixed schema, schema precision/recall against gold fields, and setup time per corpus.

Source video

Thousand-Page Ingestion Scaling Probe

beyond-evals

Per-page cost and peak memory in a rasterize-plus-visual-parse pipeline scale sublinearly only when documents are chunked with cross-page state; below some page threshold, cross-page context loss degrades extraction accuracy.

Proof of concept

Generate synthetic documents with known ground truth at 10, 100, 1,000, and 5,000 pages, run the same visual parsing pipeline with and without overlap/cross-page state, and record where accuracy and memory break.

Measurement

Accuracy versus page count, peak memory, cost per page, and latency per page; the page count at which context-loss errors appear.

Source video

Architectural implications

3

Document understanding shifts from text-layer extraction to visual rasterization.

Before

Parse the PDF text layer and feed strings to the model.

After

Render documents visually and reason over layout, handwriting, and charts like a human reader.

Consequence

Pipeline stages must handle image rendering and visual model inference, and accuracy on complex layouts is claimed to improve.

Source video

Extraction schema is a post-parse configuration artifact, not a pre-committed contract.

Before

Define a fixed schema before seeing document contents.

After

Generate or enhance the schema based on parsed document contents before extraction.

Consequence

Developers must plan for a schema-generation step in the workflow rather than hardcoding target shapes.

Source video

Ingestion becomes a composable primitive set (Parse/Split/Extract/Edit/Classify) rather than one parser.

Before

Write a monolithic custom parser per document type.

After

Compose discrete document primitives into per-product pipelines.

Consequence

Teams can build downstream products (resume parsing, flight-history extraction) by recombining stages instead of rebuilding parsers.

Source video

Tradeoffs and failure modes

3

Visual rasterization vs. text-layer extraction

Benefit

Handles rotated documents, charts, scans, and handwriting directly, reportedly improving accuracy on complex enterprise layouts.

Cost or risk

Requires rendering and visual model inference, and the summary does not quantify the compute, latency, or cost of this approach relative to naive extraction.

Significantly improves accuracy on complex enterprise layouts compared to naive text extraction.
Open source video
Source video

Agentic OCR self-correction loop

Benefit

Flagging uncertainty and correcting step-by-step reportedly reduces hallucinations and structural errors.

Cost or risk

A multi-step correcting loop implies additional passes and latency versus a single-pass transcription; the summary offers no throughput or latency figures.

Rather than making a single pass and committing to an answer, the model flags uncertainty, corrects itself, and proceeds step-by-step.
Open source video
Source video

Standard vs. enhanced schema generation

Benefit

Enhanced schema generation is tailored to parsed document contents for nuanced extractions; standard generation is quicker.

Cost or risk

The summary gives no criterion for when the extra tailoring is worth it or how correctness of generated schemas is validated.

Reducto offers both standard schema generation and enhanced schema generation tailored to parsed document contents.
Open source video
Source video

Open questions

4

How does the agentic OCR decide when to flag uncertainty and trigger self-correction?

Why unresolved

The summary states the behavior exists but does not describe the uncertainty signal, correction criteria, or stopping condition.

Research direction

Prototype an extraction loop with an explicit uncertainty signal and measure whether self-correction improves structural accuracy versus cost.

Source video

What are the latency and cost benchmarks for hybrid visual-OCR model pipelines?

Why unresolved

The summary lists this as an open question and provides no performance figures.

Research direction

Benchmark rasterization-plus-visual-model inference against text-layer extraction on accuracy, latency, and cost per page.

Source video

How does Reducto handle extremely large documents with thousands of pages at scale?

Why unresolved

The summary lists this as an open question and does not describe chunking, parallelism, or state management.

Research direction

Test scaling behavior on thousand-page documents to identify where per-page cost or memory becomes a constraint.

Source video

How is the correctness of generated (standard or enhanced) extraction schemas validated before downstream use?

Why unresolved

The summary does not describe any evaluation or review step for the generated schemas.

Research direction

Design an evaluation harness that scores extractions against gold schemas to compare standard vs. enhanced generation.

Source video

Key claims

7
factualVerification needed

Reducto parsed 3M+ pages of scanned PDFs, handwritten notes, and government files behind JMail.

Evidence

Reducto parsed 3M+ pages of scanned PDFs, handwritten notes, and government files behind JMail.

Question

What was the end-to-end time and cost to parse the 3M+ pages, and under what hardware configuration?

Source video
factualVerification needed

Reducto has processed over 2 billion documents across enterprise customers.

Evidence

Over $108M raised in funding, processing over 2 billion documents across major customers.

Question

What counts as a 'document' in the 2 billion figure and over what time period was it accumulated?

Source video
factualVerification needed

JMail was cloned from Gmail in five hours following the release of the Epstein files.

Evidence

JMail was originally built by Riley Walls and Luke Igel in five hours last November.

Question

Did the five hours include the document parsing, or only the application build?

Source video
comparativeVerification needed

Visual rasterization significantly improves accuracy on complex enterprise layouts compared to naive text extraction.

Evidence

Significantly improves accuracy on complex enterprise layouts compared to naive text extraction.

Question

On which document types and benchmarks is the accuracy gap measured, and by how much?

Source video
causalVerification needed

An agentic OCR loop that flags uncertainty and self-corrects reduces hallucinations and structural errors in extracted data.

Evidence

Reduces hallucinations and structural errors in extracted data.

Question

What is the measured reduction in hallucination or structural error rate versus single-pass OCR?

Source video
opinionVerification not requested

Enterprise document ingestion requires dedicated pre-processing layers (layout analysis, OCR correction, schema mapping) before passing data to LLMs.

Evidence

Enterprise document ingestion requires dedicated pre-processing layers (layout analysis, OCR correction, schema mapping) before passing data to LLMs.

Source video
factualVerification needed

Reducto has raised over $108M in funding.

Evidence

Over $108M raised in funding, processing over 2 billion documents across major customers.

Question

Is the $108M figure total funding raised and as of what date?

Source video

Connections

4