The Missing Infrastructure for AI Agents: Documents | Reducto’s Adit Abraham

The episode opens with an overview of the core problem: AI agents can reason over vast amounts of data, but much of the world's work—especially in sectors like healthcare, legal, finance, and insurance—is still stored in unstructured formats such as PDFs, spreadsheets, and scans. Adit Abraham emphasizes that traditional OCR methods are insufficient for these complex, messy documents. He explains that while models now aim to process these documents effectively, existing tools often fail silently on long or intricate texts, leading to critical information drops. The conversation sets the stage by highlighting the importance of reliable document processing infrastructure as a foundation for AI to perform high-stakes work accurately.

Adit Abraham discusses Reducto’s strategy of combining vision-language models with specialized infrastructure to improve accuracy, speed, and cost-efficiency in document parsing. He describes how their R1 model was developed from scratch, built after extensive research and iteration, and tailored specifically for challenging document types. The model aims to unify visual understanding and language comprehension, achieving high accuracy without the traditional trade-offs. The importance of rigorous benchmarking, real-world data challenges, and fallback strategies like self-review mechanisms is stressed as key to reaching near-perfect accuracy for high-stakes applications.

The conversation shifts to the complexities of real-world documents, which often contain dense, interconnected, and multimodal content such as dense tables, layered images, and nested PDFs. Abraham highlights that these densities and embedded multimedia make it difficult for models to faithfully interpret documents. He shares insights into specific technical obstacles, like handling multi-page convoluted tables and reconciliation of conflicting data across multiple sources. Reducto’s focus on active structuring, specialized pipelines, and dense visual understanding aims to overcome these hurdles, ensuring models process even highly complex documents reliably.

A key theme is the relationship between accuracy improvements and the expansion of AI use cases. Abraham notes that current OCR accuracy is often below 80% for complex documents, which is insufficient for automation where mistakes impact real-world decisions. They have achieved over 90% accuracy, enabling near full automation in many scenarios. Abraham emphasizes that agents capable of reviewing and correcting their work can outperform humans on structured extraction. Cost efficiency is also discussed, with Reducto's pipeline reaching a penny per page, opening the door to scaling document processing to billions of pages and unlocking vast new applications.

The episode delves into how Reducto’s technology is being adopted by major enterprises across sectors like hedge funds, legal, and healthcare, driven by demands for faster, more reliable data extraction. Abraham discusses the importance of evaluation frameworks, including benchmarking on diverse, real-world data, and the challenge of existing benchmarks that can be inconsistent or misleading. He explains the significance of flexibility—clients often have their own data sets—and how Reducto collaborates to ensure robust testing. The scalability of their processing—billions of pages monthly—demonstrates the rapid growth and broadening scope of AI-driven document workflows.

Looking ahead, Abraham envisions Reducto expanding beyond reading and parsing to enabling AI agents that can actively edit, translate, and reason over documents. He emphasizes that their focus on infrastructure and domain-specific models keeps them ahead, as large labs tend to pursue broader, less targeted AI capabilities. Reducto plans to roll out new tools, including an integrated editing plugin, and aims to make document workflows more accessible and reliable. He stresses the importance of context, verification, and active correction to trust AI in critical sectors, and describes their goal of making high-volume, complex document processing seamless, thus freeing humans from mundane tasks.