Hello,
I’m Elad.

At Rubicon Robotics, I owned the constraint enforcement layer for an ML analytics engine that turns underwater swim video into 49 interdependent performance metrics.

I modeled dependencies so simulations produced physically achievable swimmer states, audited residual formulas against the production Rust pipeline, identified errors in 15 of 21 formulas, and built a Jacobian-based elasticity framework to trace error propagation and prioritize fixes.

At the Humor Project, I lead product strategy for an AI memory architecture and work at the intersection of language, product, and machine learning.

I also built an iOS application that runs on-device ML models for real-time text generation, shaping both the product experience and the underlying technical approach.

I built no-gears-typing for the moment you realize you typed a whole sentence in the wrong language. It detects the mismatch and corrects the text as you type.

The app uses Rust and Python with an n-gram language detector that reaches 80% accuracy. It corrects keyboard-language mismatches in real time, and I’m actively developing it.

A model can claim tool-calling support and still fail with the tools you actually use. I built Callprobe to test your own tool definitions against OpenAI-compatible endpoints.

The harness scores tool selection, schema validity, and argument accuracy separately. It also tests when a model should avoid calling a tool, and how performance changes as more tools enter the conversation. The goal is to make the failures understandable, beyond a single benchmark score.