# Cooking with Reducto’s CEO: Why PDFs Still Break AI — Adit Abraham

## Executive summary

Adit Abraham, CEO of Reducto, discusses the challenges of reliable AI document parsing, arguing that general-purpose LLMs often fail on complex structures like tables and handwriting. He details Reducto's approach, which involves specialized pipelines and model routing to achieve high fidelity. Key technical advancements include the proprietary R1 model, designed for end-to-end processing and improved cost efficiency, and a focus on continuous evaluation to catch production failure modes, especially in low-margin-of-error industries like healthcare. The discussion also covers the architectural shift toward treating document processing as a search and context synthesis problem, rather than just a generation task.

## Key takeaways

- The Limitations of General-Purpose LLMs for Documents: Frontier models can fail by silently dropping rows or missing context in long documents, a failure mode not visible in production. Specialized platforms are necessary to audit and ensure data integrity, especially when 99% accuracy is insufficient for critical use cases. (10:57)
- Reducto's Architecture and R1 Model: Reducto utilizes a model routing strategy, combining smaller traditional CV models for baselines with larger models for complex tasks. The proprietary R1 model aims to provide end-to-end processing, improving accuracy and simplifying pricing by eliminating complex multipliers. (4:19, 9:29)
- Focus on Continuous Evaluation and Failure Modes: Reliability requires continuous evaluations (Evals) to spot imperfect, out-of-distribution failure patterns, not just incorrect data points. For high-stakes industries, traceability (knowing the source of extracted data) and accurate confidence scoring are critical. (13:05)
- The Nature of Document Processing as a Search Problem: The core problem is often viewed as a search problem: retrieving relevant context from multiple sources (e.g., a file system) and synthesizing it, rather than solely relying on Retrieval Augmented Generation (RAG) or simple document parsing. (12:35)

## Technical details

- Document Processing Evolution: The field has evolved from OCR to IDP (Intelligent Document Processing) and now incorporates LLMs and agents. The goal is to replicate human-level fidelity when reading documents, which is a complex vision problem. (3:35)
- Reducto's Capabilities: Reducto aims to be the best platform for any interaction with unstructured files, covering data extraction, classification, and editing. It handles the entire workflow from reading to writing into documents. (5:29)
- Model Strategy and Ownership: The company's mandate is to be the best provider regardless of whether the model is in-house or external. The stack has shifted toward in-house models to maintain control and performance consistency. (7:00)
- Scaling and Performance Metrics: Reducto processes over a billion pages monthly. Performance improvements are measured not just by accuracy, but by cost efficiency and the ability to drive the Pareto frontier (better intelligence per dollar). (9:29, 14:30)

## Practical implications

- For enterprises handling sensitive documents (e.g., healthcare, finance), relying solely on general-purpose LLMs is risky due to potential silent data loss or context omission.
- Building robust document pipelines requires specialized, proprietary models (like R1) that manage model routing and ensure end-to-end reliability.
- The focus must shift from merely achieving high accuracy to ensuring full traceability and implementing continuous evaluation loops to detect subtle, out-of-distribution failure patterns in production.
- The future of white-collar work involves not just extraction, but reliable editing (e.g., redlining contracts, translating content while maintaining layout integrity).

## Topics

AI Document Parsing, LLM Reliability, Build Engineering, Data Pipelines, Model Architecture, Scalability, Continuous Integration/Evaluation, Reducto, Deep Extract

Source: https://www.youtube.com/watch?v=KgqWyVVH4Oc
