
随着企业越来越多地转向 AI 智能体以实现工作自动化,锁定在复杂文档中的海量信息仍是一大挑战。Unsiloed AI 在 Felicis 领投下完成了 750 万美元种子轮融资,旨在通过专门技术帮助企业让文档更易于被 AI 智能体访问。
As businesses increasingly turn to AI agents to automate work, the vast amounts of information locked in complex documents remain a challenge. Unsiloed AI has raised $7.5 million in seed funding led by Felicis to tackle that problem, aiming to help enterprises make their documents more accessible to AI agents through specialized technology.
硅谷风投机构 Felicis 领投本轮融资,Y Combinator、Orange Collective 和 Uncorrelated Ventures 跟投。天使投资人 Aarthi Ramamurthy 和 Beth Turner 也参与了投资。
Silicon Valley venture capital firm Felicis led the round, with participation from Y Combinator, Orange Collective, and Uncorrelated Ventures. Angel investors Aarthi Ramamurthy and Beth Turner also participated.
商业文档通常包含表格、图表和复杂排版,AI 系统往往难以准确解读。但如果信息提取出错,可能给企业带来高昂代价;对于依赖这些文档开展日常工作的银行、医院或律师事务所而言,影响可能尤为严重。
Business documents often include tables, charts, and complicated layouts that AI systems can struggle to interpret accurately. But getting the information wrong can lead to costly mistakes for businesses, and the impact could be severe in the case of a bank, hospital, or a law firm, as they rely on these documents for their daily work.
Unsiloed 旨在用软件解决这一问题:将复杂文档拆解,从表格、图表及其他元素中提取信息,并转化为 AI 智能体可用的结构化数据。该技术还会标明提取信息在原始文档中的来源位置,以便企业核对结果,而非盲目接受 AI 的输出。
Unsiloed aims to solve this problem with software that breaks down complex documents, extracts information from tables, charts and other elements, and turns it into structured data that AI agents can use. The technology also shows where the extracted information came from in the original document, so businesses can check the results rather than take the AI’s output at face value.
“大语言模型的问题在于它们本质上是概率性的,”Unsiloed 联合创始人兼首席执行官 Aman Mishra(上图右)在采访中表示。“必然会产生某种偏差,必然会出现某种幻觉,这对下游影响巨大。”Mishra 与 Adnan Abbas(上图左)共同创立了 Unsiloed,两人是印度理工学院(IIT)Kharagpur 校区的前同学。在花费一段时间开发技术后,两人加入了 Y Combinator 2025 年秋季批次。
“The problem with LLMs is they’re inherently probabilistic,” said Unsiloed co-founder and CEO Aman Mishra (pictured above, right) in an interview. “There is going to be some deviation induced. There is going to be some hallucination involved, and that has a massive negative impact downstream.”Mishra co-founded Unsiloed with Adnan Abbas (pictured above, left), his former classmate at the Indian Institute of Technology (IIT) Kharagpur. The two joined Y Combinator’s Fall 2025 batch after spending some time developing the technology.
Unsiloed 已在金融、法律服务和医疗保健领域与客户合作,Mishra 告诉 TechCrunch。他说,一家银行正在使用其技术处理文档,而一家律师事务所已在自己的服务器上部署该技术,用于处理合同和旧案卷,但他未透露这两家客户的名称。不过,Mishra 点名了语音 AI 初创公司 Bolna AI,称其使用 Unsiloed 摄取知识库以构建语音代理。
Unsiloed is already working with customers in finance, legal services, and healthcare, Mishra told TechCrunch. One bank is using its technology to process documents, while a law firm has deployed it on its own servers to handle contracts and old case files, he said, without naming either customer. Mishra, however, named voice AI startup Bolna AI as a customer, saying it uses Unsiloed to ingest knowledge bases for building voice agents.
这家总部位于旧金山的初创公司进入了一个已经拥挤的市场,其中包括 Reducto 和 Unstructured 等专业文档处理初创公司,以及微软和亚马逊的产品。企业有时也会使用通用 AI 模型从文档中提取信息,这给 Unsiloed 增添了另一批竞争对手。
The San Francisco-headquartered startup has entered a market that is already crowded and includes specialized document-processing startups such as Reducto and Unstructured, as well as offerings from Microsoft and Amazon. Businesses also sometimes use general-purpose AI models to extract information from documents, giving Unsiloed another set of competitors.
Mishra 告诉 TechCrunch,Unsiloed 的优势在于其处理复杂文档的方式。该初创公司使用其所谓的“双流”视觉模型,分别处理文本和图像数据,并结合一种称为“热力图分块”的技术,帮助识别文档中最重要的部分。
Mishra told TechCrunch that Unsiloed’s advantage lies in how it processes complex documents. The startup uses what it calls “dual-stream” vision models, which handle text and image data separately, along with a technique called “heat map-based chunking” that helps identify the most important parts of a document.
“在一个流中你有文本标记,在另一个流中你有精确的图像标记,”Mishra 说。“这是我们花费大量时间进行架构设计的东西。”该技术随后将更多算力导向图形、数字和图表等区域。Unsiloed 还使用专用解码器,考虑行业特定术语,帮助其区分在不同业务语境下可能含义不同的术语。
“In one of the streams you have the text tokens, in the other stream you have the exact image tokens,” Mishra said. “This is something that we have spent a lot of time architecting.”The technology then directs more computing power toward areas such as figures, numbers, and charts. Unsiloed also uses a specialized decoder that takes industry-specific terminology into account, helping it distinguish between terms that can mean different things in different business contexts.
米什拉举了一项测试来支持这些说法,该测试涉及一家企业客户提供的542份财务文件。根据这家初创公司分享的一份技术简报,Unsiloed在大约8.4万个提取数据点上的准确率达到99.78%。该测试将Unsiloed与主流AI模型、微软的Azure文档智能以及该客户内部的提取工具进行了对比。不过,这些结果尚未经过独立验证。
Supporting those claims, Mishra pointed to a test involving 542 financial documents supplied by an enterprise customer. Unsiloed achieved 99.78% accuracy across roughly 84,000 extracted data points, per a technical brief shared by the startup. The test compared Unsiloed with major AI models, Microsoft’s Azure Document Intelligence, and the customer’s in-house extraction tool. The results have, however, not been independently verified.
Unsiloed还声称,在Allen AI开发的公开基准测试olmOCR-Bench上,其表现优于竞争的AI文档处理工具。在这家初创公司进行的测试中,Unsiloed得分为88.0,而Reducto的代理系统为66.0,Unstructured为39.9。该基准测试覆盖1403页PDF,考察了这些系统处理表格、数学表达式、扫描文档和复杂版面的能力。
Unsiloed also claims to have outperformed competing AI document-processing tools on olmOCR-Bench, a public benchmark developed by Allen AI. In tests conducted by the startup, Unsiloed scored 88.0, compared with 66.0 for Reducto’s agentic system and 39.9 for Unstructured. The benchmark covered 1,403 PDF pages and tested how well the systems handled tables, mathematical expressions, scanned documents, and complex layouts.
米什拉表示,Unsiloed的定价为每页1美分起。他声称,这家初创公司的方法还有助于降低成本,因为它将算力集中投向文档中最需要关注的部分,而不是对每个部分都采用同样的处理方式。
Unsiloed’s pricing, Mishra said, starts at one cent per page. He asserted that the startup’s approach also helps keep its costs down by directing computing power toward the parts of a document that need the most attention, instead of processing every section in the same way.
目前,Unsiloed有七名员工。米什拉表示,这家初创公司计划利用新筹集的资金招聘更多研究人员和工程师,扩大销售力度,并开发专有数据集以改进其技术。
Currently, Unsiloed has a headcount of seven people. The startup, Mishra said, plans to use the fresh funding to hire more researchers and engineers, expand its sales efforts and develop proprietary datasets to improve its technology.