NYTech
People
Personfounder

George Sivulka

George Sivulka is the founder and chief executive of Hebbia, an artificial intelligence company based in New York that builds agent-based software for financial and legal research and other white-collar knowledge work [1][2][3]. He holds a PhD in electrical engineering, a master's degree in applied physics, and a bachelor's degree in mathematics, all from Stanford University [4][5][6]. Sivulka has described himself as one of the youngest students admitted to Stanford and the youngest member of his PhD cohort, and he was pursuing fully-funded doctoral research on metalearning, the idea of building machines that can learn to generalize across tasks, when he founded the company [10].

Sivulka traces Hebbia's origin to his time teaching large introductory math classes at Stanford, where he observed that many of his most talented students went on to jobs at firms such as Morgan Stanley and Goldman Sachs and appeared miserable performing repetitive analytical work [9]. Around 2016 and 2017 he began noticing early transformer models and advances such as improved Google Translate and neural information retrieval, which convinced him a major shift was coming in how documents could be searched and analyzed [9]. He built an early neural information-retrieval tool in a Jupyter notebook and shared it with former students working in finance so they could analyze lengthy filings such as DEF 14A disclosures; the tool spread informally among people he had never met [9]. The catalyst for actually leaving his PhD program came in mid-2020, when OpenAI released GPT-3 under a paper describing large language models as metalearners, which he took as evidence that the technology he had hoped to build himself already existed and that his effort should shift to applying it [8][10].

According to Sivulka, Hebbia's product evolved from a search-oriented tool into a broader platform that lets non-technical users direct large language models over documents to produce outputs such as extraction pipelines, chatbots, and structured data, a shift he compares to how Excel took a technical capability (the SQL database) and made it usable by ordinary professionals [9]. He argues that new technologies require the right product to become meaningful, drawing an analogy to fire needing the torch and the wheel needing the chariot, and contends that chat-style interfaces are only an early, "single-threaded" way of presenting AI rather than its final form [8]. A central technical claim he makes is that Hebbia reduces hallucination by first identifying supporting citations within source documents and only then generating text, rather than generating text and searching for justification afterward [8]. He also describes the company's early architecture around a spreadsheet-like interface in which each cell functions as an independent agent operating over data, which he presents as a first step toward orchestrating multiple collaborating agents on complex, multi-step tasks [8].

Sivulka states that Hebbia's client base has included large asset managers, banks, law firms, and government clients, and that the company has avoided per-client fine-tuning of its models, aiming instead for a general-purpose platform in the manner of Microsoft Word or Excel [9]. He has said the company raised early backing from investors including Peter Thiel, Jerry Yang, and Ram Shriram, with additional funding from Index Ventures, and has cited later fundraising totaling around 160 million dollars [9][10]. He has described rapid growth in usage, saying the platform processed roughly 100 million pages of unstructured data in a prior year and was on pace to process four to five billion pages in the following year [10]. Sivulka has also said the company grew to roughly 100 employees working primarily out of a New York office, with expansion into San Francisco and London and a target of 300 to 400 employees by year's end [10].

Founded

Experience

  1. Founder & CEO
    Hebbia.AIAug 2020 to Present

Education

Media & appearances

In the news

This page shows public professional information only, each fact cited. Is this you? Corrections or removal within 24 hours, no questions asked.