Felicis 领投 750 万美元,Unsiloed AI 让企业文档可被智能体使用

随着企业日益转向 AI 智能体来实现工作自动化,锁在复杂文档中的海量信息仍是一大难题。Unsiloed AI 已获得由 Felicis 领投的 750 万美元种子轮融资,以解决这一问题,目标是借助专门技术,帮助企业让其文档更易于被 AI 智能体获取。
硅谷风投公司 Felicis 领投本轮融资,Y Combinator、Orange Collective 和 Uncorrelated Ventures 参投。天使投资人 Aarthi Ramamurthy 和 Beth Turner 也参与了投资。
商业文档往往包含表格、图表和复杂版式,AI 系统可能难以准确解读。但信息出错会给企业带来代价高昂的错误,而在银行、医院或律师事务所这类依赖这些文档开展日常工作的机构中,影响可能尤为严重。
Unsiloed 希望通过软件解决这一问题:该软件将复杂文档拆解,从表格、图表及其他元素中提取信息,并将其转化为 AI 智能体可用的结构化数据。该技术还会标明所提取信息在原始文档中的来源位置,因此企业可以核对结果,而不是直接采信 AI 的输出。
“LLM 的问题在于它们本质上是概率性的,”Unsiloed 联合创始人兼 CEO Aman Mishra(上图右一)在接受采访时说。“必然会产生一定的偏差,也会涉及一定的幻觉,而这会在下游造成巨大的负面影响。”
Mishra 与 Adnan Abbas(上图左侧)共同创立了 Unsiloed,后者是他在印度理工学院(IIT)卡拉格普尔分校的同学。两人在花了一段时间开发这项技术后,加入了 Y Combinator 2025 年秋季批次。
Mishra 向 TechCrunch 表示,Unsiloed 已经在金融、法律服务和医疗健康领域与客户展开合作。他说,一家银行正在使用其技术处理文档,一家律师事务所则将其部署在自有服务器上,用于处理合同和旧案件文件,但他没有透露这两家客户的名称。不过,Mishra 点名语音 AI 初创公司 Bolna AI 是其客户,称该公司使用 Unsiloed 摄取知识库,以构建语音智能体。
这家总部位于旧金山的初创公司进入的是一个已经相当拥挤的市场,其中既有 Reducto、Unstructured 等专注文档处理的初创公司,也有 Microsoft 和 Amazon 提供的产品。企业有时还会使用通用 AI 模型从文档中提取信息,这又给 Unsiloed 带来了一批竞争对手。
Mishra 向 TechCrunch 表示,Unsiloed 的优势在于其处理复杂文档的方式。这家初创公司采用其所谓的“双流”视觉模型,将文本和图像数据分开处理,并结合一种名为“基于热力图的切块”的技术,帮助识别文档中最重要的部分。
“在其中一个流中是文本 Token,在另一个流中是精确的图像 Token,”Mishra 说。“这是我们花了大量时间进行架构设计的地方。”
随后,该技术会把更多算力投向图表、数字和图形等内容。Unsiloed 还使用了一个专门的解码器,将行业特定术语纳入考量,帮助它区分在不同业务语境下含义不同的术语。
为支持这些说法,Mishra 提到了一项测试,涉及某企业客户提供的 542 份财务文档。根据这家初创公司分享的一份技术简报,Unsiloed 在大约 84,000 个提取数据点上实现了 99.78% 的准确率。该测试将 Unsiloed 与多家主流 AI 模型、微软的 Azure Document Intelligence 以及该客户自研的提取工具进行了对比。不过,这些结果尚未经过独立验证。
Unsiloed 还声称,在 Allen AI 开发的公开基准测试 olmOCR-Bench 上,其表现超过了竞争的 AI 文档处理工具。在这家初创公司开展的测试中,Unsiloed 得分为 88.0,而 Reducto 的智能体系统为 66.0,Unstructured 为 39.9。该基准测试覆盖 1,403 页 PDF,考察这些系统处理表格、数学表达式、扫描文档和复杂版面的能力。
Mishra 表示,Unsiloed 的定价为每页 1 美分起。他称,这家初创公司的方法还通过把算力集中投向文档中最需要关注的部分,而不是对每一部分都以同样方式处理,帮助其压低成本。
目前,Unsiloed 共有 7 名员工。Mishra 表示,这家初创公司计划利用这笔新融资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。
当您通过我们文章中的链接进行购买时,我们可能会赚取少量佣金。这不会影响我们的编辑独立性。
Jagmeet 为 TechCrunch 报道印度的初创企业、科技政策相关动态以及所有其他重大科技进展。他此前曾在 NDTV 担任首席记者。
您可通过发送邮件至 [email protected] 联系 Jagmeet,或核实其主动联络的真实性。
正在加载下一篇文章
译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。
查看英文原文(Felicis leads $7.5M bet on Unsiloed AI to make enterprise documents usable by AI agents)
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.
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.
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 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.
“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 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.
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 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.
“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.
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 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.
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.
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.
When you purchase through links in our articles, we may earn a small commission. This doesn’t affect our editorial independence.

Jagmeet covers startups, tech policy-related updates, and all other major tech-centric developments from India for TechCrunch. He previously worked as a principal correspondent at NDTV.
You can contact or verify outreach from Jagmeet by emailing [email protected].
Loading the next article
原文为英文,正文由自动抓取并保留原文。
原文发布于 TechCrunch AI(Felicis leads $7.5M bet on Unsiloed AI to make enterprise documents usable by AI agents),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。