Tushar Madan went from measuring pitchers to measuring whether AI agents can be trusted

Before the agent boom, Tushar Madan spent his time on a question baseball has been arguing about for a century: how do you actually measure whether a pitcher is any good? His published Databricks work took the Minnesota Twins' pitch scenario analysis and scaled it with Spark and R, turning a mountain of individual pitches into something a coaching staff could make decisions from. Performance measurement under uncertainty, at volume, with real money riding on the answer.

Hold that thought, because it explains everything about what he is doing now.

The unfashionable questions

In 2026 roughly everyone is building AI agents. Almost nobody is asking whether the agents are worth what they cost, what happens when they quietly fail, or who is accountable when one acts on your behalf and gets it wrong. Avianna, the independent research lab Tushar co-founded with Rishubh Khurana, exists to ask precisely those three questions: agent economics, agent reliability, alignment and governance.

That framing is sharper than it sounds. The industry's default demo shows an agent completing a task; it does not show the failure rate, the recovery path, or the bill. Avianna's bet is that the interesting engineering of the next few years is not making agents more capable, it is making them accountable enough to hand real work to. Enterprise software has been here before with every prior wave, and the companies that survived were the ones that treated reliability as the product.

The output is not just papers. Concord is a contracts layer for agent actions, so what an agent is allowed to do is written down rather than vibed. Lattice is the worker layer of the governance stack. Both are attempts to give agents the thing every other production system already has: rules, and a way to check them.

And then there is LeetMath

The third project is the tell. LeetMath is probability training for adults, a practice track for the exact reasoning skill most people, including most people shipping AI, are quietly bad at. A lab studying whether machines reason reliably decided that humans could use the drills too, which is either a sly joke or the most consistent thing on the list. We think both.

Avianna advises a small number of startups and enterprise teams, deliberately small, which is the correct scale for a research lab that wants to keep saying true things. If you are shipping agents and have not yet asked what happens on the day one of them is confidently wrong, start at avianna.ai. The pitching data suggests he has been thinking about failure rates for a while.