Xperiment

Hassen Poreya

Staff GTM data scientist in the San Francisco Bay Area.

I work out where marketing and product money actually pays off, then build the measurement so the answer holds up when someone pushes on it. About eight years across attribution, experimentation and forecasting, now as the founding data scientist at Agaro AI.

Now
Founding data scientist at Agaro AI, since October 2025
Leading
Two data engineers, as people manager and technical lead
Before
Adobe, Intuit and Mercury Insurance
Focus
Multi-touch attribution, marketing mix models, A/B testing and causal inference, GTM measurement

The work, by the evidence

Every highlight below is tagged by company, theme and tool. Click a company on the timeline, a dot in the grid, or a tool to filter the list.

Tools in the highlights

Methods

Regression-based marketing mix models, multi-touch attribution, A/B and holdout testing, propensity score matching, difference-in-differences, matched cohorts, funnel and cohort analysis, LTV and CAC modeling, XGBoost, Prophet forecasting.

AI toolkit

Claude Code routines and subagents, Claude Tag in Slack, supervised-autonomy agent workflows, Cursor.

    How I read an experiment

    One Creative Cloud test at Adobe, walked through the way I would in a review: personalized template recommendations, measured on whether people went on to use the product.

    1. 1

      Pick the metric first

      Primary metric was 7-day downstream activation. Click-through was rejected as a vanity metric before anything launched.

    2. 2

      Size it honestly

      User-level randomization on about 15% of traffic for roughly two weeks, powered to detect a 2% lift.

    3. 3

      Read it by segment

      The topline moved about 3%, but not evenly. Pick a segment on the right to see what each one showed.

    4. 4

      Decide, then fix what it exposed

      Ship to the segment where it worked, fix the desktop instrumentation, and re-test new users.

    Directional summaries of the test, not a full readout.

    What I am building now

    At Agaro AI

    I built the data function from zero and lead it: attribution, mix modeling and automated reporting that every team uses weekly, plus Claude skills that answer routine data questions in Slack and an automated A/B loop with human approval gates.

    Xperiment, on the side

    Synthetic AI agents that simulate user behavior on web and UI flows, so low-traffic products can still run experiments. It is in progress, and you can try the early version.

    Run an A/B test with synthetic agents

    Education

    B.S., Computer Science

    Herat University

    Foundations of Data Science

    Stanford Continuing Studies, 2019