Hey HN! I'm Aleks Misztal, one of the cofounders behind tracelight.
We’ve built an AI that understands spreadsheet financial models and it's available directly in Excel. You can chat with your whole spreadsheet and get natural language descriptions of the precedents in every formula.
Behind the scenes, it's like a coding agent - it combines reasoning models with tools such as semantic search, formula evaluation, and precedent tracing. It's already pretty good (took a lot of iteration on prompts and tool input and output schemas) and will get much better over the coming weeks.
We've also trained a specialised model for semantic understanding of spreadsheets. It's able to assign meaning to each cell in the sheet based on the labels, layout, formulas, styles, etc. We use it to power semantic search and to give feedback to the agent when it uses other tools.
You can download it for free on the Excel add-in store. Let us know what you think!
Hey, no api yet. Generally our users find it very good for help with short range tasks, especially exploratory ones, where error tolerance is higher. For api use cases usually the reliability requirements are higher. We did build a way to extract data from many spreadsheets at once for one of our enterprise pilots. For another, we are experimenting with running this with a human in the loop so that we can give higher reliability guarantees while still getting the speed up from AI.
Curious what's your use case?
health economic models built for pharma. You often have one base model and then have to generate multiple country adaptations, which also require updates to a report (which can be ~100 pages). Right now, this is a manual task for which consultancies charge up to $100K+. Reliability is indeed crucial though, but human in the loop might be fine.
We’ve built an AI that understands spreadsheet financial models and it's available directly in Excel. You can chat with your whole spreadsheet and get natural language descriptions of the precedents in every formula.
Behind the scenes, it's like a coding agent - it combines reasoning models with tools such as semantic search, formula evaluation, and precedent tracing. It's already pretty good (took a lot of iteration on prompts and tool input and output schemas) and will get much better over the coming weeks.
We've also trained a specialised model for semantic understanding of spreadsheets. It's able to assign meaning to each cell in the sheet based on the labels, layout, formulas, styles, etc. We use it to power semantic search and to give feedback to the agent when it uses other tools.
You can download it for free on the Excel add-in store. Let us know what you think!