About Yuval

Yuval Yatskan is the CEO of ePlane AI, with more than 20 years in SaaS behind him, most recently in cybersecurity. The company started in 2016 as a marketplace for the aviation ecosystem and pivoted around 2023 to building AI products on top of the unusual dataset it had accumulated. Its newest product, AeroGenie, is an agentic decision engine that reaches well past aviation — and, at the time of recording, was weeks away from launch.

What we talked about

From marketplace to decision engine — why sitting on a unique dataset was worth more than running the marketplace built around it.

Simulations versus machine learning — why you can skip training a model entirely, and what you get instead of a single point estimate.

Where the language model actually sits — his case that chatbots are a brilliant proof of concept, and what has to be built around them.

What makes a decision defensible — replayability, reasoning, and a hashed audit trail for the moment someone asks you to justify a call from two months ago.

The cross-industry stress tests — a legal strategy, a product price, an investment portfolio, and a 120-slide deck.

A team under ten — how the build economics changed, and why saying no to customers is now the hardest part of the job.

Sharpest moments

On the division of labour with chatbots:

LLMs are awesome, chatbots are great for what they were designed to do. But when they end, this is where we begin.

On what changes when you stop making trade-offs:

We don’t hallucinate, we actually read the entire data, join all the tables.

On the legal case a law firm had already worked for weeks:

It did find vulnerabilities we hadn’t been aware of, and it makes a lot more sense.

On what a decision needs that an answer doesn’t:

You’re asked to defend your decision. You want to make sure that the data that was used at the time that might have changed since hasn’t been lost.

On the trap of being good at everything:

Sometimes your capabilities could become your own curse.

On what has genuinely changed for founders:

Today a team of one can do so many things.

Key takeaways

• Where a chatbot samples — skipping lines, joining three tables instead of five — AeroGenie is built to ingest everything, which he attributes to writing the core in Mojo (he cites 200x to 1,000x+ over equivalent Python).
• Simulations need no trained model and no pre-chosen feature set; they return a lower bound, an upper bound, and a most likely value rather than one number.
• Simulation and machine learning are complementary — a simulation can pick the features for a model, or measure the probabilities around a model’s output.
• A decision, unlike an answer, has to be replayable and defensible: AeroGenie hashes its output (256-bit) so the call can be reconstructed with the data as it stood at the time.
• The test cases came from wildly different fields: a legal strategy the firm’s own team had missed (about a minute, roughly $1.86 in tokens), a launch price revised from $75 to $45, a portfolio stress test, and a 120-slide deck restructured by persona.
• Fewer than ten people work on the product — in his framing, the constraint has moved from engineering capacity to deciding what to build.
• The hardest discipline is saying no: profitable one-off deployments are what stop a product company from scaling.