Researching noses, brains, behavior, and data (Ph.D., Weizmann).
Chemical signals
Human chemical communication
Ph.D. research · Weizmann Institute of Science
Isolated a single-molecule human body-volatile chemosignal with no perceived odor
(hexadecanal) and demonstrated its behavioral, physiological, and brain-activity effects,
modulating human aggression with opposite-sign effects by sex.
Noise-blast level chosen in each of 27 rounds. Left to right: men control, men HEX, women
control, women HEX. Darker is louder. n = 130.
paperSci. Adv. 2021, Fig. 2B·dataMendeley, TAP_FullData.csv
A $12 sensor, a naturalistic output
The field measured aggression with a button press: pressed or not pressed. I built a
pressure sensor that works inside an MRI scanner: participants clench to express how
strongly they act, so aggression is recorded as a continuous, naturalistic output. It
captures the behavior better than the button press, and it cost
$12 against the $2,000
off-the-shelf part. It is
open source, and the study it
produced ran in Science Advances.
Humans sniff themselves
We asked whether people are constantly, and mostly unaware, smelling their own bodies:
how often the hand goes to the face, and to the nose, in the lab and in a survey across
countries. Co-first author; the paper made the issue cover.
Subjective sleep quality from a consumer smartwatch
Whenever the outcome is a human experience, the person belongs in the model.
Subjective sleep quality is the thing clinicians actually care about and the thing no
device measures: a watch records movement and heart rate, not whether you feel rested.
We built a framework that combines continuous wearable physiology (sleep architecture,
activity, HRV-derived stress) with static sociodemographic and personality features, and
infers the rating at the individual level.
Trained on the PerMed cohort: 4,794 participants,
759K daily questionnaires,
1.77M wearable-days over three years. End-to-end, from raw
signal through modeling.
What the model listens to. SHAP values for its 8 strongest features, 500 reports each.
Darker is a higher feature value.
paperunder review·dataanonymized release, forthcoming
Passive and continuous assessment of subjective sleep quality using smartwatches.
Hausner, O.*, Mishor, E.*, Reiner, M., Shmueli, E. Under review, Communications
Medicine. *equal contribution.
Co-founded and co-built an AI-native decision-support platform that applies scientific
reasoning to early-stage company building. Led product development end-to-end: from
founder interviews, product strategy, and reasoning-framework design to implementation,
evaluation, and launch.
Open-ended AI output can’t be judged by feel. Find a ground truth the model never sees,
then check against it deterministically.
One row per company, one mark per documented failure mode: caught (light to dark, easy
to hard) or missed (dot). 61 of 78 caught. Mean score 4.57 of 5 across five criteria;
weakest on non-obvious insight (4.2).
paperinternal benchmark·data15 public post-mortems, private repo
In progress
Code-review ledger. Every finding from AI code reviewers and its disposition. The research plan is done; analysis running.
Prompt pressure. Does emotional force with no new information change a model's code-review verdict? Runs in progress.