Building Tracelight

Today we’re launching Tracelight. Tracelight is an AI company developing better ways for businesses to build and use quantitative models, and we’re starting by solving the interface between spreadsheets and LLMs.

When businesses make their most consequential decisions, they build models. For the most part, these models are mechanical frameworks that map how different elements of a business or market interact and how specific actions lead to specific outcomes. They serve as both analytical tools and coordination mechanisms: they make assumptions explicit, create a shared language for debate, and they inform and reflect the legal contracts that govern any resulting agreement. They underpin everything from M&A, to capital raising, to strategic planning.

We are in the middle of an intelligence revolution, but these high-stakes business decisions - where, arguably, intelligence has highest leverage - remain largely untouched. Transactions and strategic decisions don't happen faster than before, the process isn't cheaper, and there's little evidence that they are higher-quality. That’s because there are a lot of hurdles in the way. We're starting with a big one: spreadsheets.

Today, financial models are built in Excel. Spreadsheets remain the most powerful way to create and share quantitative thinking, and billions of them power the world’s economy. While humans remain the decision-makers, spreadsheets aren’t going anywhere. But they are fundamentally different from the text, images, and structured data that LLMs typically handle well. They combine formulas, values, formatting, and spatial relationships in ways that make them impossible to simply feed into ChatGPT. For now we are stuck with hacky workarounds like copying and pasting regions as markdown, or Microsoft’s Copilot’s ‘select a rectangle of data and press a button to tell Microsoft that it’s a table’. Users hate it.

A few months ago, we picked up the latest research on representing spreadsheets to LLMs, and we’ve been improving on it since. Today, we encode arbitrary spreadsheets into compressed linguistic representations that LLMs can work with. Tracelight understands a spreadsheet like a human would: it reads the values, formulas, formats, precedents and dependents of cells, and it considers the spatial relationships between cells. This is the key to unleashing today’s AI into spreadsheets; we’re very excited by the progress we have made on it.

And Tracelight is already adding value. We are helping alternative asset managers and transaction consultants to build financial models that are error-free and more easily handled downstream. Our goal with pilot customers is to reduce by 90% the time required to work with these models, while improving accuracy with AI-powered error-checking and smart human validation workflows. For decision-makers, this will make a big difference: better work, in less time, at lower cost. In time, we think Tracelight can become the product that enterprises use to power all their biggest decisions.

We’re thankful for the support of our pre-seed investors, Entrepreneur First and Inovo, and we’re very excited by the potential applications of our work. If you think Tracelight could help you, or you’re an engineer looking for ambitious work, please message us on founders@tracelight.ai.