An inside look at Ramp Labs
You might be familiar with Ramp Sheets, an AI spreadsheet editor that went viral on X when it launched in November.
Ramp Sheets is part of a series of products launched by Ramp Labs, a group focused on AI experiments within Ramp’s Applied AI team.
Last month, Ramp Labs’ own Alexander Shevchenko and Will Koh joined 8VC in San Francisco for Chat8VC, their monthly series that gives technologists a platform to showcase their work. Shevchenko and Koh chatted about what Ramp Labs is, why it was formed, and demoed Ramp Sheets.
Here are key insights from the conversation.
The experimental team with no mandate
Ramp Labs runs experiments with AI, both in applied AI and in areas such as reinforcement learning. The team was created to explore AI without predetermined problems to solve.
“Oftentimes the really interesting use cases only come up when you're doing pure explorations in the tech itself,” Shevchenko summed it up.
Since there is no mandate, the team has a lot of freedom to explore whatever technology they’re interested in.
“There's no direct metric that we optimize for, like impressions or conversions from what we put out, or reusability of the tech," Shevchenko explained.
“Obviously, we try to tie it back into Ramp in some way.”
Alexander Shevchenko and Will Koh chat with 8VC's Vivek Gopalan
Ramp Sheets
Ramp Sheets is a great example of what’s possible with this exploratory approach.
The project started as a process mining effort to help Ramp's accounting team with documentation work related to the month-end close. Process mining, Shevchenko explained, is a branch of data science that analyzes event logs from complex systems to build out process maps, find bottlenecks, and suggest optimizations.
"This project was completely unrelated to Ramp Sheets," Shevchenko said. "But only through explorations in the tech did we weasel our way into using these very applied LLM agents for spreadsheet traversal."
Through trial and error with regular prompting, the team explored how an LLM can think about spreadsheets. Shevchenko elaborated on the team’s thinking: How can we teach an agent to navigate multiple sheets and understand complex formulas spread across them so they can create a wide range of tasks for users? What actions should it take? How much context should it have?
They refined this work through what Shevchenko called “vibe evals,” running early outputs by Ramp’s finance and product teams who are familiar with Excel’s best practices and the application’s full capabilities. They iterated until the output met the standard finance users sought.
The result was Ramp Sheets, an AI-native spreadsheet editor to build models, clean data, write formulas, and even search the web. Designed like ‘Cursor for Excel,’ the tool has the standard Excel interface with an editor but with the AI’s reasoning made visible to the user. “It has all the common formulas you'd expect, but it also shouts out all the information there as you would expect in Cursor,” Shevchenko explained.
Under the hood, Ramp Sheets is powered by an OpenAI Agent SDK connected to a frontier LLM. An Excel sheet evaluation service performs actions such as applying or dragging formulas, rereading the output, and agentically sending over the output to the frontend for recalculation and display on the Excel editor. The tool also has access to a code interpreter for tasks where it can't calculate something in Excel itself.
Although Ramp Sheets was born from freestyle exploration, the team has discovered specific use cases for it, especially in bookkeeping, forecasting, modeling, and error detection. Outside of Ramp, X is replete with examples of how people are using the tool — from generating slide decks and charts, to building operating models for valuations and dashboards, to scouting vendors for wedding planning.
Shevchenko believes that this is the first iteration of vibe finance the world is moving towards. "The goal is to get finance people very comfortable with using LLMs and AI assistance to get their work going."
Ongoing experiments
Besides Ramp Sheets, the team is experimenting in areas such as generative user interfaces (UIs). Generative UIs adapt dynamically to user sessions to produce context-aware experiences, rather than static ones. Shevchenko explained, "Maybe having the same UI through hard coding every path and the long tail of possibilities doesn't make the most sense every time.” So, the team is experimenting with building widgets and smaller sections in the website that are generated on the spot. If you're looking at a vendor page on the Ramp website and have a pending transaction, a widget could appear on the page and let you approve the transaction immediately.
Another area of interest for the team is video-based process mining, which applies the concept of process mining to video rather than to event logs. The system could analyze a Loom recording of someone processing invoices or reconciling accounts and use it to produce a process map. To Shevchenko, this is intuitive: “Whenever you're talking to a designer or an engineer, you send them a Slack message and often they'll respond with a Loom.” He added that video offers much more context than text.
Video-based process mining has not been productionized yet by any startup, which makes this a promising research frontier. Ramp Labs is experimenting with models like Gemini 3 to turn videos into structured outputs: “You throw in a video, you ask for a process map or you ask how to find bottlenecks in it,” Shevchenko said.
The team is also experimenting with reinforcement learning to make spreadsheet traversal faster and more cost-efficient (see their early work using Tinker).
Hiring for high agency
Ramp’s Applied AI is still a young, 17-person team and is growing fast.
“We're looking for people who have a lot of autonomy and ownership,” Koh emphasized. “We want people who believe that there's no problem out there that you can't solve these days. People who can execute and come back with results that show their progress.”
Koh elaborated that there’s a lot of ways at Ramp to apply AI to make experiences better and faster, and take care of tedious tasks for customers.
