AI Engineer · LLM agents and evaluation · Netanya, Israel
AI agents that do research. Tests that prove what they found.
I build multi-agent systems, and the evaluations that separate real discovery from a model repeating what it has already read.
01
Theseus Fly
Can an AI scientist discover a mechanism, or only recall one?
Language models have read most of what is known about the fruit fly brain, so a correct answer proves little. Theseus Fly asks the question on brains that do not exist anywhere else: rewired copies of the real connectome, running in a full brain simulation, with the answers sealed before the agents begin.
PILOT RESULT · SCORE ON SEALED ANSWERS · p = 0.01
A model that only remembers the literature scores 0.132. The distance above that line is discovery, not memory.
02
Eidolum
A ledger that checks every prediction.
Analysts make thousands of public market calls and almost nobody keeps score. Eidolum does, automatically, every day. I built it and run it alone, from the data pipelines to the LLM that judges the cases rules cannot.
03
Method
How a result gets onto this page.
Decide before looking. The test and the bar it must clear are written down before anything runs.
Audit my own work. Independent leak checks and rebuilds that must match bit for bit.
Keep the failures. Negative results stay in the record, and every claim names where it holds.
Verify the agents too. I build with fleets of coding agents and trust none of their output until it passes the same gates.
04
Who I can help
Where I can be useful.
Data teams
I built the data pipeline behind Eidolum and I run it alone. Four licensed market data sources come in every day, get checked and land in PostgreSQL, over 577,000 records so far. I like the unglamorous part of the job: catching bad data before it reaches the numbers people see.
Teams building agents
I work with dozens of coding agents in parallel every day, and I have an LLM making real decisions inside a product people use. I can help you ship agents and put tests around them, so you hear about it when they go wrong and not from your users.
AI evaluation teams
Most of my work comes back to one question: can this answer be trusted? I design tests with sealed answers, honest baselines and leak checks, and I write down what failed as carefully as what worked.
NeuroAI labs
I reproduced a published whole fly brain model bit for bit and built the tooling that lets AI agents run experiments on it. If your lab has a model or a dataset and needs an engineer who can turn it into experiments you can rely on, I would like to hear about it.
05
About
Trained on the mat first.
Years of competing taught me the habits I bring to research: train every day, test yourself against the best, and let the result decide. I study for a B.Sc. in Intelligent Systems Applications at the Open University of Israel.
BRAZILIAN JIU JITSU
Israeli national team · coach and seminar instructor
Building agents, or need to know if yours can be trusted?
Open to AI Engineer and AI evaluation roles in Israel or remote, and to research collaborations.