Hai Rozencwajg

Hai Rozencwajg

Founding Team | Member of Technical Staff @ Accomplish

Actions

Hai Rozencwajg is a Member of Technical Staff and part of the founding team at Accomplish, with a decade of experience in data science and machine learning. His work at Accomplish focuses on building and evaluating self-learning AI workflows, with an emphasis on skill quality, agent memory, and model routing. He is a co-author of Boundary-Bench, an open-source benchmark for evaluating AI coding agents in hardened environments (NeurIPS 2026 workshop). Hai holds B.Sc. and M.Sc. degrees in Computer Science from Bar-Ilan University. Passionate about generative AI, data, and coding, he loves learning new things and mentoring others. He is married to Almog and has three children.

We Tried OSS Model Routing Tools So You Don’t Have To

Modern AI applications are no longer single prompts. They are pipelines of decisions: retrieval, reasoning, tool use, and execution. However, most systems still default to a single model across all steps, leading to unnecessary cost, latency, and failure modes.

This talk presents a hands-on evaluation of open-source model routing tools in the context of multi-step AI workflows. Instead of comparing models in isolation, we examine how routing decisions change when facing a multi-step pipeline as part of computer-use task execution.

We’ll cover real experiments, architectural patterns, and common pitfalls, and provide a practical approach to step-aware routing that improves both performance and cost efficiency.

Agent vs Machine: How Production Boundaries Change Agent Behavior

We’ve all seen an agent get stuck doing something absurd: retrying the same command, looking for another package source, attempting sudo, changing the environment, or burning thousands of tokens trying to get around a restriction.

Sometimes that behavior isn’t just the model being bad at the task. The environment changed the task.

Production agents operate behind network policies, filesystem boundaries, non-root users, permission systems, and sandbox controls. Once one of those boundaries fires, it becomes an input into the agent’s execution loop: the agent may give up, retry, find a valid workaround, take a much longer path, or repeatedly fight a restriction it can never overcome.

In this talk, we'll share what we learned from creating Boundary Bench: we ran the same agent tasks under different production-like constraints and analyzed not only whether agents succeeded, but how their execution changed along the way.

We’ll look at real agent trajectories, how different models respond to the same restriction, and why pass/fail alone misses some of the most useful information for anyone operating agents in production.

Hai Rozencwajg

Founding Team | Member of Technical Staff @ Accomplish

Actions

Please note that Sessionize is not responsible for the accuracy or validity of the data provided by speakers. If you suspect this profile to be fake or spam, please let us know.

Jump to top