Applied Scientist, Behavior Modeling

Why Shuttlebase?

Every day, millions of people decide whether to buy something online, and the reasons behind those decisions are still mostly guesswork. We want to change that.

We're working to quantify human psychology: to build an embedding model for human behavior that understands how people react to a page, a price, or a single sentence, and can predict what will work before anyone runs a test.

Every experiment our AI runs on a live store teaches that model something about how people decide. Our goal is for any company to be able to ask, "I'm selling product X to audience Y, how do I optimize for Z?" and get an answer it can trust.

We're early and well funded. The people who join now will pick the tools, design the models, and help choose who joins next. If that's a problem you want to spend the next few years on, we'd like to meet you.

Every day our script records tens of millions of actions from people visiting the websites we work with: what they look at, where they hesitate, what they click, and when they leave. You'll turn that into the first version of our behavior model. The goal is a model that predicts how people are likely to interact with a website, and what the site should and shouldn't do for them.

We'll start with online stores, since that's where we work today, by spotting which visitors are close to buying and which groups need to see a different version of the page. Researchers have published a lot about learning from user behavior with sequence models and embeddings, so you won't start from a blank page. Your job is to know that work well, pick what fits our data, get it working here, and keep up as the field moves.

You'd be the first person working on this full time, so a lot rests on your shoulders!

• The behavior model. Build our first model that turns raw session events into embeddings of how a visitor behaves. Pre-training a transformer on event sequences is a likely start, but you'll make the call. The hard part is that most sessions are short and anonymous, and most visitors never come back.

• Buying intent and audiences. Use the model to spot buying intent during a session and to group visitors into audiences, so the AI can show each audience the page version that works for it. We want a first version on live stores as soon as it works, and we'll improve it from there.

• New research. When a paper comes out that could help, you'll be the one who reads it, tries it on our data, and decides whether we use it. If you learn something worth publishing, publish it.

We mostly work in Python, with event data in Click House and services on AWS. You'll report directly to our CTO, and work every day with our engineers, who build on your model, and with our client team, who explain to stores why the AI showed their shoppers a different page. A lot of what they say will come from you.

• You read papers and get impatient until you've tried the idea on real data.

• You'd rather adapt a proven method and see it work in three weeks than spend a year inventing a new one.

• You want a big, open problem that's yours to figure out, with engineers and the founders working right next to you.

• You're fine working with noisy event logs and writing your own SQL to get at them.

• You were the person in your lab others came to when they needed a hard idea made simple.

• You want to help build a model that understands how people make decisions, and you'd like to be there for the first version of it.

• A PhD or equivalent research experience in a quantitative field. People from physics, math, computational biology, and neuroscience often fit well, especially if you've modeled behavior or sequences.

• You've trained a model on…