George Sivulka built Hebbia after former Stanford Math 51 students at Morgan Stanley and Goldman Sachs were stuck with repetitive document work, then took a leave and never returned.
Sivulka argues Hebbia can do for large language models what Excel did for the SQL database, letting ordinary users apply AI to documents without technical training.
Sivulka warns that chaining steps that are each 90% accurate can yield an output only about 10% good, so Hebbia is targeting near-100% accuracy per step, citing 99.95%.
Hebbia wants to be the Excel of language models
Hebbia is betting that LLMs need a spreadsheet-style tool that non-technical knowledge workers can use on their own documents. Finance is the starting wedge because analysts spend their days on repetitive document questions. The hard part is reliability across multi-step tasks, where small per-step errors compound, so the product competes on accuracy and trust rather than on the base model.
Takeaways
Hebbia aims to be the Excel of LLMs. Sivulka calls Hebbia the first LLM-native productivity tool. Excel brought the SQL database to ordinary users and, he says, took over 90% of fixed income within about eighteen months. Hebbia wants to give non-technical users the same reach with language models.
Private equity due diligence was the beachhead. Analysts fill out 100- or 400-question trackers from data rooms to decide on investments worth hundreds of millions. Sivulka says no private equity CEO discussed an LLM strategy six months ago, and now many large North American and European asset managers are customers.
Multi-step tasks are where generative AI breaks. Models handle first-order tasks but lack context and repeated reasoning for nth-order ones. Chaining steps that are each 90% accurate can yield output only about 10% good. Hebbia targets near-perfect steps and cites 99.95% accuracy.
The moat is input, output and flexibility. Input covers integrations, document parsing and multiple indexes. Output means steering models programmatically toward user-controlled results. Flexibility makes results understandable and trustworthy. Sivulka wants to stay ahead of better base models, citing OpenAI's function-calling API as overlapping startup work.
The product works across models without customization. Hebbia grew from neural search into a suite of LLM tools for ETL pipelines, search, chatbots and data extraction. Customers choose their own models. Sivulka says it serves large asset managers, banks and the US government without fine-tuning for any customer.
Humans stay central and no jobs are replaced. Sivulka compares Hebbia to spreadsheets replacing ledgers and says no job has been replaced so far. Making humans better is his stated goal, and he notes users do not trust LLM output by default.
Expansion beyond finance is underway. Sivulka jokes Hebbia would have failed if it stayed in financial services. It has begun serving law firms and is growing its government work, with the ambition of being as basic a tool as Chrome or Excel.
How Hebbia started
Sivulka, a PhD student and computational neuroscientist, was put in charge of a Stanford Math 51 class with many hundreds of students.
Tap a name to go there · tap a line to see why they are linked
On this episode 2
Show 1
Many students became knowledge workers at Morgan Stanley or Goldman Sachs, where he saw them unhappy with repetitive tasks.
Early transformer models around 2016 and 2017 impressed him.
He built a Jupyter notebook running an early neural information retrieval model.
Former students in Morgan Stanley's Menlo Park office used it on the roughly 400-page DefM14A document.
He took a leave of absence from his research and never went back. Hebbia was founded in 2020.