Hai Rozencwajg

Hi, I’m Hai.

Data scientist. Building AI agents. Evaluating… Vibing…

I’m a data scientist on the founding team at Accomplish. These days, I work on how AI agents learn from experience, and what happens when they get stuck.

Portrait of Hai Rozencwajg
Hai Rozencwajg
Based in Tel Aviv

01 / Selected work

What I’m
working on.

Building / Founding team

Accomplish

I’m a Member of Technical Staff in Data Science on the founding team. I build and evaluate self-learning AI workflows.

I’m interested in whether an agent can use what happened on one task to do better on the next.

Research / Co-creator

Boundary-Bench

I co-created Boundary-Bench to study coding agents under production-like restrictions.

We run the same tasks under different policies to see how restrictions change success rates, cost, and the ways agents fail.

Personal project / Chrome extension

Gmail Task Indicator

I built a Chrome extension that adds blue checkmarks to Gmail threads linked to Google Tasks, so email follow-ups are easier to spot.

I like breaking work into manageable tasks and keeping track of what needs doing. This is one small tool I built around that habit.

02 / Research

Research
I’ve shared.

03 / Talks

Talks I’ve given,
and what’s next.

Upcoming

Agent vs Machine: How Production Boundaries Change Agent Behavior

AiDevTLV · With Niv Yungelson

Recording

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

AWS × DevOps · With Niv Yungelson

Talk

From Email Overload to Actionable Insights

NLP IL × Taboola

04 / About

A bit
about me.

I’ve spent about a decade working in data science and machine learning across real estate, cybersecurity, and ad-tech, at startups and larger companies. I’m now part of the founding team at Accomplish, working on AI agents and how they learn.

Earlier, I was part of the founding team at Skyline AI, which was acquired by JLL in 2021.

I have a BSc and MSc in Computer Science from Bar-Ilan University. My studies focused on machine learning and network security.

I also have two patent applications from my time at JLL, covering novel entity resolution using generative AI and domain-specific recommender systems. Entity resolution means matching records that refer to the same real-world entity. The recommender work accounts for geospatial context, company hierarchies, and other domain signals.

I’m married to Almog, and we have three children who are getting into AI too. They’ve even been inventing their own games. I enjoy learning new things and helping others learn, too.

If you’re working on AI agents, evaluations, or memory, I’d be happy to hear what you’re learning.

Say hello