Key Moments
Cooking with Reducto’s CEO: Why PDFs Still Break AI — Adit Abraham
Latent Space PodcastKey Moments
AI struggles to parse PDFs accurately, losing context in tables and complex layouts, demanding specialized solutions beyond general LLMs, even as companies like Reducto develop advanced internal models.
Key Insights
Reducto processes over a billion pages per month, processing over 100 million pages daily, and has grown its sales team from 1 to 30 people in a year.
PDFs were designed for print accuracy, not machine readability, posing a complex visual problem akin to self-driving car research.
Reducto's new R1 model aims for end-to-end processing, including layout capture and final-stage corrections, offering better accuracy at a lower cost than basic models.
While frontier models can miss rows and lose context, Reducto's Deep Extract tool reportedly outperforms them, highlighting the need for specialized document parsing.
Reducto is evolving from document extraction and analysis to editing, aiming to empower models to reliably review contracts and fill out forms.
The company's hiring process prioritizes passion and a "world-class" level of obsession with a chosen field, assessing how candidates approach challenges and continuous improvement.
The inherent complexity of PDF parsing for AI
The fundamental challenge in AI document parsing, as discussed by Adit Abraham, CEO of Reducto, lies in the nature of PDFs themselves. Originally designed for print accuracy, PDFs were never intended for machine readability. This creates a "very complex visual problem" for AI, much like the challenges faced in developing self-driving cars. Unlike simple text extraction, parsing PDFs requires understanding visual layouts, tables, and nuanced formatting. Companies have been tackling this for decades, but the context of document interaction has drastically changed, necessitating AI that can read and process documents with human-like accuracy, which is a significant hurdle when the source format wasn't built with AI in mind.
Reducto's evolution from memory to document intelligence
Reducto's journey began with a pivot from their initial YC-backed company, Remember, which focused on long-term memory for AI agents. The name 'Reducto' emerged organically through a brainstorming session for new names and domain availability. Initially exploring concepts like 'Long Horizon Memory' or 'Memory for Agents,' the company has now solidified its focus on document intelligence. This evolution reflects a strategic shift towards addressing a critical need in the AI landscape: the ability to reliably process and understand unstructured data within documents, moving beyond simple data extraction to more complex analysis and manipulation.
The limitations of frontier models and the rise of specialized solutions
Adit Abraham highlights a critical limitation of large language models (LLMs) and frontier models: their struggle with complex document structures like tables. While these models are powerful, they can miss entire rows or lose crucial context in long documents, a failure mode that is difficult for users to detect. Reducto's 'Deep Extract' tool, for example, is designed to address these shortcomings, reportedly outperforming general models in accuracy. This emphasizes that for specific, high-stakes tasks like document parsing, specialized solutions are often necessary to achieve the required precision and reliability, rather than relying solely on broad, general-purpose AI models.
Introducing Reducto R1: An end-to-end document processing model
Reducto is set to release a new internal model called R1, designed to handle document processing from start to finish. This model integrates components for layout capture and final-stage corrections, aiming for near-perfect accuracy. The development of R1 signifies a move towards proprietary, highly optimized models that can outperform foundational models across dimensions like accuracy, cost, and latency for specific tasks. The goal is to provide a seamless and highly efficient experience for users, abstracting away the complexities of underlying AI models and focusing on delivering exceptional results.
Bridging the gap from extraction to editing and action
Reducto's vision extends beyond merely extracting and analyzing information from documents. They aim to become the premier platform for any interaction between AI models and unstructured documents, encompassing reading, analysis, classification, and crucially, editing. This expansion into document editing addresses a significant gap in current AI capabilities. The ability for AI to not only comprehend but also reliably modify documents—like reviewing contracts or filling out loan applications—opens up a vast array of new use cases. This move is seen as a 'Claude Secret Reveal moment' for office work, akin to the transformative impact of LLMs on code generation.
Navigating evaluation and ensuring reliability in low-error domains
Ensuring accuracy and reliability in document parsing is paramount, especially for industries like healthcare and finance with extremely low error tolerance. Reducto focuses not just on extraction accuracy but also on traceability and the trustworthiness of human review processes. They measure how reliably users report issues, understanding that even 99% accuracy might not suffice if critical edge cases are missed. This involves rigorous internal testing, continuous evaluation of model performance against real-world usage patterns, and identifying specific failure modes, such as difficulties with handwritten notes on angled pages, to inform future training.
The future of Reducto: Horizontal platform and AI-driven editing
Reducto envisions a future as a horizontal platform, offering broad capabilities rather than vertical, industry-specific solutions. Their focus remains on the core document processing actions—extraction, analysis, and editing—regardless of the industry. While document storage is not a primary focus, the company aims to be the world's best at the actual procedures performed on documents. This includes developing sophisticated editing capabilities that can handle complex formatting changes, language translations with layout adjustments, and enabling AI to perform tasks previously done by humans, such as filling out loan applications or reviewing legal contracts with high reliability. The recent launch of a model editing endpoint is a significant step in this direction, with more features planned to revolutionize document-centric workflows.
Building a culture of passion and continuous improvement
Reducto's hiring philosophy centers on identifying individuals with a deep 'passion' for their work, characterized by meticulous attention to detail and a drive for continuous improvement. They look for a 'world-class' obsession with a chosen field, whether it's AI research, cooking, or sports. This is assessed through interviews that probe into what candidates are passionate about and how they approach challenges. The company culture fosters an environment where engineers are encouraged to explore and adopt new AI tools organically, demonstrating value through tangible results rather than mandates. This approach aims to build a team that is not only technically proficient but also deeply invested in mastering their craft and solving complex problems.
Mentioned in This Episode
●Supplements
●Software & Apps
●Companies
●Organizations
●People Referenced
Common Questions
Reducto is a company developing AI-powered tools to address the challenges of processing, analyzing, and editing unstructured documents like PDFs. They aim to make interactions between AI models and documents more seamless and accurate, improving upon traditional OCR and IDP methods.
Topics
Mentioned in this video
A company focused on AI-powered document processing, analysis, and editing.
A company whose early Claude model was used in a hackathon that contributed to Reducto's founding idea.
The previous company Adit Abraham was involved with when joining Y Combinator.
A self-driving car company, used as an analogy for the complexity of achieving full autonomous capability, similar to Reducto's goal with document processing.
A company with which Reducto is collaborating to create a measurement benchmark.
A startup accelerator program where Adit Abraham's previous company, Remember, joined.
An earlier iteration or name related to Reducto's development, focusing on memory for agents.
A company Reducto works with, providing legal services, highlighting sector-specific expertise.
A widely known AI chatbot, mentioned in the context of the co-founder's familiarity with AI systems.
A tool that can reportedly create presentations more efficiently than manual effort.
A Reducto tool for document analysis that reportedly outperforms many foundational models.
A system that Reducto relies on for accuracy, used in benchmarking discussions.
A communication platform used within Reducto, mentioned in the context of founder availability and team communication.
A previous Reducto model based on Kwen VLM, used for experimentation.
An AI model from Anthropic, mentioned in relation to a hackathon where Reducto's founding idea emerged.
Current AI models used in document processing and agent interactions.
A tool or platform mentioned in comparison to Reducto's capabilities, related to document editing and knowledge work.
A vision-language model that Roll MOCR was based on.
A tool that Reducto's sales team is extensively using to create custom websites for prospects.
A leading AI startup that Reducto serves, indicating the competitive nature of the market.
An early large language model, mentioned in relation to a highly capable young engineer named Arth.
More from Latent Space
View all 275 summaries
51 minMini Brains, Beating Heart Organoids, and the AI That Predicts Them — Fabian Theis, Helmholtz Munich
73 minBehind Latent Space, ft AIEi Shanghai preview — with swyx and Galen Gao
35 min⚡️ The 2026 Software Factory Infra Stack — Zach Lloyd, Warp.dev
68 minBuilding Legal Agents When You Can't See the Data — Jacob Lauritzen, CTO of Legora
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free