访谈
Adit Abraham, " 降级访谈 " 联合创始人兼首席执行官系列
在 Google 上的首选来源中添加 Unite.AI
阿迪特·亚伯拉罕(Adit Abraham)是Drapto公司的共同创始人和首席执行官，他专注于将复杂的企业文件转化为AI应用的可靠数据。 在与麻省理工学院研究生劳纳克·乔杜里(Raunak Chowdhuri)创办公司之前，他曾在麻省理工学院学习计算机科学，在麻省理工学院媒体实验室进行机器学习研究，并担任Google产品经理的Ads and Search工作。 当不可靠的文件摄入成为反复出现的障碍时，创建者努力为大型语言模型建立内存工具。 亚伯拉罕与早期的客户的直接工作帮助塑造了公司对准确性，安全性，部署灵活性和可预见生产性能的强调。
Drapto开发文件智能基础设施，帮助AI团队将PDF，扫描，电子表格等文件转换为结构化，可使用的信息。 它的平台结合了光学字符识别，计算机视觉，和视觉语言模型来解释文本，表格，数字，格式化，以及复杂的页面布局。 除了解析外，它还提供了结构化取出、文件分拆、分类和编辑的工具，以及将取出的信息与来源相连接的引用。 客户包括Harvey，缩放AI，万塔，默科和罗高。 公司的工作涵盖金融、保健和法律等领域的检索、检索和文件工作流程，而准确的摄取和可追踪的结果至关重要。
Drapto源于你和Raunak Chowdhuri在为大型语言模型建立记忆工具时遇到的挑战。 与现有的文件处理工具有什么特别的突破，你何时意识到解析问题本身已经足够大，足以在周围建立一个公司？
我们试图让模型从文档中访问上下文，所以我们增加了一个文件上传功能，但我们不断遇到处理质量的问题。 当时我们尝试了所有现有的工具，但是当它们都不符合我们的标准时，涉及到真实世界的混乱文件，人们实际上上传。
因此，我们建立了我们自己的解析解决方案，最初是作为我们LLM内存产品的一个特征，但后来从开发者那里看到了惊人的拉动，他们自己也会遇到这个问题，并且热爱我们的解决方案。 有些团队花费数周时间建造和维护自己的摄取管道，这些对话使我们更清楚地看到这一机会，因为尽管模型智能越来越好和更好，但清洁和准确的背景仍然大有不同。
PDF技术设计是为了显示信息，而不是使信息可以机器读取。 是什么使得复杂的文档，特别是那些包含表格，图表，多栏布局，或嵌入式图像的文档，如此难以AI系统可靠地理解？
当我们作为人类阅读一个文件时，我们把许多我们自然的理解视为理所当然的 — — 我们可以几乎立即推断出两者之间的关系，因为我们可以看到一个数字属于某个专栏标题下，这个文本在另一个专栏中继续，或者一个标题指的是上面的图。
AI模型实际上确实与它相冲突，因为PDF并不是用来保留这类关系信息或元数据；它们基本上是非常数据密集的图像。 一个解析器必须从页面中重建这些关系，而问题在合并的单元格，嵌入式表格，图表，扫描和异常布局上变得更加棘手。 因为我们专注于这个文件理解问题，我们可以可靠地处理AI模型往往产生幻觉并搞砸的长尾文件。
将计算机视觉、光学字符识别和视觉语言模型纳入其文件处理程序。 这些不同的技术如何共同运作，传统的OCR系统在哪里仍然往往不足？
传统的OCR已经持续了数十年，它是为了逐个读取包含干净，清晰的文字的文档而建造的。 在介绍手写、扫描、扭曲或表格文件时，这很容易分崩离析，这些文件往往占企业数据的大多数。
简化为开创性地将视觉语言模型添加到“Agentic OCR ” 的文档处理空间中，因为我们看到了VLM不仅仅理解文档中纯文本的潜力 — — 真正的文档智能需要背景、准确度和多模式能力。 我们最新的 r-1 解析模型结合了 Agentic OCR 的功率具有格式化，引出，布局检测等特性成一个单通。
这一额外工作很重要，因为即使一个整体精确度很高的解析器，如果它漏掉的少量信息包含某个人实际需要的数字、条款或表格条目，也会造成严重的问题。
随着AI模式日益多式并能够直接解释图像和文件，你如何看待专门文件智能平台的作用演变？基础模型能否最终完成大部分这项工作，或者专用基础设施是否仍有必要？
事实上，我们相信AI模型只会改进并变得更快和更好 — — 随着新模型不断涌出，我们不断更新评估它们的弱点，并将它们的长处融入我们自己的管道，为我们客户提供总体最好的性能。 帕斯多想做你的摄入团队帮助你成长
企业仍然面临一个独特的挑战，即不仅需要业绩，而且需要在生产规模上实现可扩展的吞吐量、治理和可核查性，这对于今天的基础模型来说仍然是困难的。
最近，我们用GDP.pdf 基准测试测试了这一点，它显示前沿模型在复杂的真正专业文件中仍然难以解释，在严格的任务级别通过率上得分低于15%至25%。 通过让Prapto先分析这些文件，然后将结构化的数据传递到前沿模型，宏观精确度提高了9个百分点，同时与直接消耗文件的模型相比，推理符号减少了13%。
这种显示对于专用摄取基础设施仍有巨大的价值，因为大多数企业数据仍然不饱和。
许多企业AI应用程序最终依赖于检索的生成或代理系统访问内部文件所埋藏的信息。 这些系统中的可靠性问题究竟有多少是从文件摄取和分析开始的？
在摄入过程中开始大量 — — 由于坏答案出现在输油管的尽头，各小组往往在很晚的时间才发现问题。 如果一个解析器从表格中掉出一行，混合了两个列的读取顺序或者错过了一个图，则检索无法恢复甚至从未进入数据库的信息。
我们通常看到，随着RAG和搜索系统更一致地返回，我们的客户终端用户在将摄入量与Placto交换后，其经验大有改善。 例如，Harvey看到其与OCR有关的客户票大幅下降，在作为加工层实施Plaxto之后，产品使用量增加。
Drapto已扩展，不再局限于结构化取出和更广泛的文件工作流程。 一旦一个系统能够准确理解一个文件的内在结构，下一层文件智能看起来像什么？
数据摄取通常只是更大业务流程中的第一步。 一家公司可能收到一份500页的文件，需要查明其中的文件，对其进行分类，提取具体信息，将该信息同另一个系统进行比较，并将不寻常的案件送交专家审查。 我们花了两年时间在市场上培训最好的文件理解模型，而现在我们花更多的时间专注于如何精简之后的事物。
一旦你能够可靠地理解一个文件的内容和结构，你就可以使更多的工作自动化，目前人们花时间在系统之间组装、检查和移动。 几乎每个高考业务工作流程都可以被抽象为步骤，其中既包括人类阅读、决策，也包括写入其他文件 — — 我们正期待成为人类或agents未来与文件数据互动的默认方式。
您拥有的客户从AI本土公司，如Harvey，万塔，Scale AI,Mercor和Rogo，到主要金融机构和"财富"十大公司。 AI启动的文件处理要求与大型受监管企业有何不同？
AI-本土公司往往会快速移动，他们的工程团队想要一个API，在不强迫他们自己建立和维护处理堆栈的情况下处理困难文件。 大型企业关心的是相同的准确性，但是它们经常在数据存在的地方，如何部署软件，数量猛增时会发生什么，用户能否将提取出的价值追溯回原始页面上。 例如，规模AI在保密和预先部署重要的政府环境中使用Drapto，而Harvey则在大约六周内开始生产之前，对我们进行了精确、可靠、成本、耐久和失败模式的评价。 与这两种类型的客户合作，促使我们既注重在实际生产要求下如何减少排放，也注重其处理单个文件的准确性。
据报道，在建立一个专门的销售组织之前，主要通过创始人主导的销售活动，Pacutto达到了500多万美元的年度经常性收入。 与那些早期的客户直接交谈的是什么？教你关于开发者真正试图解决的问题？
这些对话教导我们少花时间广泛谈论文件处理问题，多花时间了解客户真正想要实现的目标。 早期的销售电话常常变成工作电话，顾客分享中断管道的文件，在医疗保健方面，这意味着要看到一个被丢弃的桌排或实验室报告中的错误读取单元如何会拖延索赔或留下一个病史不完整的临床医生。 这些桩子重新设置了我们的栏杆，以便准确性：强基准测试分数如果掩盖最重要的错误，就意味着很小。 客户需要准确性，但也需要可靠性，安全性，部署灵活性和可预见性能的规模，以及能直接影响我们所构建的事物的听觉。
目前每月处理几十亿页。 在这一规模上出现了哪些工程挑战，特别是围绕不同文件类型的延迟性、准确性、成本和保持一致的结果？
我已失去我们处理的网页数，给每一页添加不必要的推论会增加成本和耐用性，同时对速度进行过度优化会损害到困难文件的准确性，这意味着我们花费了大量的工程精力来确定每一页的处理实际需要多少。
随着r-1的出现，我们提高了解析精度，同时降低了延迟度，使价格达到每页一分一分，我们还在努力争取更自动的路由，这样简单的一页就不会消耗与包含一个复杂表或数字的资源相同的资源。 根据我们目前的篇幅，这些决定每天影响上百万页，这使效率成为核心工程问题的一部分。
展望未来几年，随着AI agents越来越多地与合同互动，金融
本文由 AI宝库自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Interviews
Adit Abraham, Co-Founder and CEO of Reducto – Interview Series
Add Unite.AI to your preferred sources on Google
Adit Abraham is the Co-Founder and CEO of Reducto, where he focuses on turning complex enterprise documents into reliable data for AI applications. Before founding the company with fellow MIT graduate Raunak Chowdhuri, he studied computer science at MIT, conducted machine learning research at the MIT Media Lab, and worked on Ads and Search as a product manager at Google. Reducto emerged from the founders’ efforts to build memory tools for large language models, when unreliable document ingestion became a recurring obstacle. Abraham’s direct work with early customers helped shape the company’s emphasis on accuracy, security, deployment flexibility, and predictable performance in production.
Reducto develops document intelligence infrastructure that helps AI teams convert PDFs, scans, spreadsheets, and other files into structured, usable information. Its platform combines optical character recognition, computer vision, and vision-language models to interpret text, tables, figures, formatting, and complex page layouts. Alongside parsing, it offers tools for structured extraction, document splitting, classification, and editing, with citations that connect extracted information to its source. Customers include Harvey, Scale AI, Vanta, Mercor, and Rogo. The company’s work spans retrieval-augmented generation, search, and document workflows in fields such as finance, healthcare, and law, where accurate ingestion and traceable results are essential.
Reducto grew out of challenges you and Raunak Chowdhuri encountered while building memory tools for large language models. What specifically broke with existing document-processing tools, and when did you realize the parsing problem itself was large enough to build a company around?
We were trying to give models access to context from documents so we added a file uploading feature, but we kept running into problems with quality of the processing. We tried all the existing tools out there at the time, but none of them met our standards when it came to working with real-world messy documents that people actually uploaded.
So we built our own parsing solution originally as a feature of our LLM memory product, but then saw amazing pull from developers who’d run into this problem themselves as well and loved our solution. Some teams were spending weeks building and maintaining their own ingestion pipelines because the available tools couldn’t handle their documents reliably. Those conversations made the opportunity much clearer to us, because despite model intelligence getting better and better, clean and accurate context still made a huge difference.
PDFs are technically designed for displaying information rather than making that information machine-readable. What makes complex documents, particularly those containing tables, charts, multi-column layouts, or embedded images, so difficult for AI systems to understand reliably?
When we as humans read a document, we take a lot of our natural understanding for granted – we can infer relationships between almost instantly, because we can see that a number belongs under a particular column heading, that text continues in another column, or that a caption refers to the figure above it.
AI models actually really struggle with this, as PDFs aren’t designed to retain this type of relational information or metadata; they’re essentially very data-dense images. A parser has to reconstruct those relationships from the page, and the problem gets much harder with merged cells, nested tables, charts, scans and unusual layouts. Because we focus on this document understanding problem, we can reliably process the long-tail of documents that AI models tend to hallucinate and mess up on.
Reducto combines computer vision, optical character recognition, and vision-language models in its document-processing pipeline. How do these different technologies work together, and where do traditional OCR systems still tend to fall short?
Traditional OCR has been around for decades, and was built to read documents that contained clean, legible text on a character by character basis. This easily falls apart when you introduce documents that are handwritten, scanned, skewed, or tabular, which tends to be the majority of enterprise data.
Reducto pioneered adding vision language models to the document processing space with “Agentic OCR” because we saw the potential for VLMs to understand more than just the pure text in documents – real document intelligence needs context, accuracy, and multi-modal abilities. Our newest r-1 parsing model combines the powers of Agentic OCR with features like formatting, citations, and layout detection into one single pass.
That extra work matters because even a parser with very high overall accuracy can cause a serious problem when the small amount of information it misses contains the figure, clause or table entry someone actually needs.
As AI models become increasingly multimodal and capable of directly interpreting images and documents, how do you see the role of specialized document intelligence platforms evolving? Could foundation models eventually perform much of this work themselves, or will dedicated infrastructure remain necessary?
We actually are betting on the fact that AI models will only improve and become faster and better – with new models constantly coming out, we stay up to date on evaluating their weaknesses, and integrate their strengths into our own pipelines to offer the best performance overall to our customers. Reducto wants to act as your ingestion team, helping build with you as you grow.
Enterprises still have the unique challenge of needing not just performance, but scalable throughput, governance, and verifiability at production-scale which is still difficult for foundation models today.
We recently tested this with the GDP.pdf benchmark, which showed that frontier models still struggle to reason on complex real professional documents and scored under 15% to 25% on strict task-level pass rates. By letting Reducto parse these documents first and then passing the structured data to the frontier models, macro accuracy improved by nine percentage points while reducing reasoning tokens by 13 percent compared to the models consuming the documents directly.
This shows that there is still immense value to dedicated ingestion infrastructure as most enterprise data is still unsaturated.
Many enterprise AI applications ultimately depend on retrieval-augmented generation or agentic systems accessing information buried inside internal documents. How much of the reliability problem in these systems actually begins upstream with document ingestion and parsing?
A significant amount starts during ingestion – teams often discover the problem much later because the bad answer appears at the end of the pipeline. If a parser drops a row from a table, mixes the reading order of two columns or misses a figure, retrieval cannot recover information that never even entered the database.
We commonly see our customer’s end users’ experience improve dramatically after swapping their ingestion with Reducto, as RAG and search systems return better results more consistently. Harvey, for example, saw their OCR-related customer tickets drop drastically and product usage multiply after implementing Reducto as their processing layer.
Reducto has expanded beyond parsing into structured extraction and broader document workflows. What does the next layer of document intelligence look like once a system can accurately understand the underlying structure of a document?
Data ingestion usually represents just the first step inside a much larger business process. A company might receive a 500-page file and need to identify the documents inside it, classify them, extract specific information, compare that information with another system and send unusual cases to a specialist for review. We’ve spent 2 years training the best document understanding models on the market, and now we’re spending more time focusing on how to streamline what comes after.
Once you can reliably understand the contents and structure of a document, you can automate more of the work people currently spend their time assembling, checking and moving between systems. Almost every high-stakes business workflow can be abstracted into steps that involve either humans reading from, making decisions, and writing into other documents – we’re looking to become the default way in which humans or agents interact with document data in the future.
You have customers ranging from AI-native companies such as Harvey, Vanta, Scale AI, Mercor and Rogo to major financial institutions and Fortune 10 companies. How different are the document-processing requirements of an AI startup compared with a large regulated enterprise?
AI-native companies tend to move quickly, and their engineering teams want an API that handles difficult documents without forcing them to build and maintain their own processing stack. Large enterprises care about the same accuracy, but they often have additional requirements around where data lives, how they deploy the software, what happens when volume spikes and whether users can trace an extracted value back to the original page. Scale AI, for example, uses Reducto in government environments where confidentiality and on-premises deployment matter, while Harvey evaluated us across accuracy, reliability, cost, latency and failure modes before moving into production in roughly six weeks. Working with both types of customers has pushed us to focus as much on how Reducto performs under real production requirements as on how accurately it handles an individual document.
Reducto reportedly reached more than $5 million in annual recurring revenue largely through founder-led sales before building a dedicated sales organization. What did speaking directly with those early customers teach you about the problems developers were actually trying to solve?
Those conversations taught us to spend less time talking broadly about document processing and more time understanding what customers were actually trying to accomplish. Early sales calls often became working sessions where a customer shared the documents that broke their pipeline, and we’d dig into what happened downstream when extraction failed. In healthcare, that meant seeing how a dropped table row or misread unit in a lab report could delay a claim or leave a clinician with an incomplete patient history. Those stakes reset our bar for accuracy: a strong benchmark score means little if it hides the errors that matter most. Customers needed accuracy, but also reliability, security, deployment flexibility and predictable performance at scale, and hearing that directly shaped what we built.
Reducto now processes billions of pages per month. What engineering challenges emerge at that scale, particularly around latency, accuracy, cost and maintaining consistent results across very different document types?
I’ve already lost track of the number of pages we process, so even a small inefficiency can become expensive very quickly. Adding unnecessary inference to every page increases both cost and latency, while optimizing too aggressively for speed can hurt accuracy on difficult documents, which means we spend a lot of engineering effort determining how much processing each page actually requires.
With r-1, we improved parsing accuracy while reducing latency and bringing the price to one cent per page, and we’re also working toward more automatic routing so that a simple page doesn’t consume the same resources as one containing a complicated table or figure. At our current volume, those decisions affect millions of pages every day, which makes efficiency part of the core engineering problem.
Looking several years ahead, as AI agents increasingly interact with contracts, financial
Auto-fetched and translated by AI Gem Box. All rights belong to the original author.