Here’s the thing. Most enterprise AI talk today starts with chatbots and ends with glossy demos. Meanwhile, the data that actually runs a business lives in rows, columns, and time stamps. That gap is where my conversation with Vanja Josifovski, CEO of Kuma.ai, really comes alive. Vanja has spent two and a half decades helping companies turn data into decisions, from research roles at Yahoo and Google to steering product and engineering at Pinterest through its IPO and later leading Airbnb Homes. He’s now building Kuma.ai to answer an old question with a new approach: how do you get accurate, production-grade predictions from relational data without spending months crafting a bespoke model for each use case?

Vanja explains why structured business data has been underserved for years. Images and text behave nicely compared to the messy reality of multiple tables, mixed data types, and event histories. Traditional teams anticipate a prediction need, then kick off a long feature engineering and modeling process. Kuma’s Relational Foundation Model, or RFM, flips that script. Pre-trained on a large mix of public and synthetic data warehouses, it delivers task-agnostic, zero-shot predictions for problems like churn and fraud. That means you can ask the model questions directly of your data and get useful answers fast, then fine-tune for another 15 to 20 percent uplift when you’re ready to squeeze more from your full dataset.

What stood out for me is how Kuma removes the grind of manual feature creation. Vanja draws a clear parallel to computer vision’s shift years ago, when teams stopped handcrafting edge detectors and started learning from raw pixels. By learning directly from raw tables, Kuma taps the entirety of the data rather than a bundle of human-crafted summaries. The payoff shows up in the numbers customers care about, with double-digit improvements against mature, well-defended baselines and the kind of time savings that change roadmaps. One customer built sixty models in two weeks, a job that would typically span a year or more.

We also explore how this fits with the LLM moment. Vanja doesn’t position RFM as a replacement for language models. He frames it as a complement that fills an accuracy gap on tabular data where LLMs often drift. Think of RFM as part of an agentic toolbox: when an agent needs a reliable prediction from enterprise data, it can call Kuma instead of generating code, training a fresh model, or bluffing an answer. That design extends to the realities of production as well. Kuma’s fine-tuning and serving stack is built for high-QPS environments, the kind you see in recommendations and ad tech, where cost and latency matter.

The training story is another thread you’ll hear in this episode. The team began with public datasets, then leaned into synthetic data to cover scenarios that are hard to source in the wild. Synthetic generation gives them better control over distribution shifts and edge cases, which speeds iteration and makes the foundation model more broadly capable upon arrival.

If you care about measurable outcomes, this episode shows why CFOs pay attention when RFM lands. Vanja shares examples where a 20 to 30 percent lift translates into hundreds of thousands of additional monthly active users and direct revenue impact. That kind of improvement isn’t theory. It’s the difference between a model that nudges a metric and a model that moves it.

By the end, you’ll have a clear picture of what Kuma.ai is building, why relational data warrants its own foundation model, and how enterprises can move from wishful thinking to practical wins. Curious to try it yourself? Vanja also points to a sandbox where teams can load data and ask predictive questions within a notebook, then compare results against in-house models. If your AI plans keep stalling on tabular reality, this conversation offers a way forward that’s fast, accurate, and designed for the systems you already run.

From Pinterest and Airbnb to Kuma.ai: Reinventing Enterprise AI

Tech Talks Daily