PostHog Research

Training user behavior foundation models and transparently sharing the research we do

We hold one of the richest behavioral datasets anywhere — events and replays of how real software gets used. Nobody has trained a foundation model on data like this. We're the first, starting with an encoder for the raw stream behind session replay.

We share the work as we go, through papers, conferences, and published progress.

What we're researching

Four tracks, each producing papers and technical reports.

  • 01

    Training a Replay Encoder model

    In progress

    A foundation model pretrained on the raw event stream behind session replay, using a multi-axis RoPE built on additive Euler angles.

  • 02

    Training a session replay model

    In progress

    Our current replay vision system uses Gemini to watch session recordings as video. We're building a new model that doesn't rely on computer vision and has access to additional data in PostHog, like logs.

  • 03

    Training a predictive user behavior model

    In progress

    Modeling what users do next from behavioral sequences.

  • 04

    Tuning a self-driving model

    In progress

    Tuning models to observe real product usage, diagnose what is broken, and act on it, benchmarked against real product problems.

We're hiring

Come take these models from whiteboard to arXiv, with the dataset, compute, and freedom to publish.

AI research engineer

2 open roles

Train our own deep ML models on petabytes of product data and ship them into a self-driving product. This is not a Jupyter notebook job.

Remote
Learn more & apply

Want the bigger picture first? Meet the AI Research team.

Papers

Papers, preprints, and technical reports from the team.

Waltz, N. · 2024

Grafting: Making Random Forests Consistent

arXiv preprint · Published before joining PostHog ·

Waltz, N. · 2024

Time Series Clustering Using DBSCAN

arXiv preprint · Published before joining PostHog ·

The first paper from our current research is in progress and will be linked here on release.

Blog

What we learn — the wins, the errors, and the projects we cancelled.

Team

The AI Research team, and the founder who can't stay away.

Events

We talk about this work at meetups and conferences.

No research talks on the calendar right now

Check back soon, or browse everything else happening on the events page.

Frequently asked questions

Put our research into practice

Most of what we build ships as a public beta long before it's polished. See what's available to try in the app.