The TWIML AI Podcast

Zoe Weil on ranking that moves you, not finds things

Zoe Weil· Co-founder and CEO at Sequen
·~52 min·English·TWIML
AI CompanyTrainingInferenceAI InfrastructureBusiness Strategy
TL;DR

Sequen reranks retail products in real time from the signals most systems ignore, such as hover time and scroll speed, to influence behavior in a high-intent moment, judging itself by what people buy rather than what they click.

01Core Mental Model

You Can't Unsee It

For its entire history the field tried to help people find things faster; Sequen rejects that goal and instead tries to influence a shopper's behavior in the moment they already have high intent.

that is really the golden thread because once you realize it and you see, hey, even though you're operating in the search space, you're not trying to help people find things, you can't unsee it.

— Zoe Weil, The TWIML AI Podcast
Key Insight
The reframe is quiet but consequential: if success is measured by purchases rather than clicks or relevance, then relevance becomes a means rather than the goal. Weil's own example makes it concrete — she says that showing more tennis rackets after you have added one to the cart would most likely lower conversion, so the most relevant result can work against the real objective. The implication is that the field's comfortable proxy, relevance, and what retailers actually want, revenue, were never quite the same target.
02The Data Gap

The Papers Don't Transfer

The biggest advances in ranking are published by social media companies, but a Wayfair or Priceline scientist who implements those papers gets a fraction of the lift, because the method was tuned to data they will never have.

It doesn't drive the same outcomes for you as it does for them because you don't have the data that they have.

— Zoe Weil, The TWIML AI Podcast
Key Insight
This is a quiet indictment of how applied ML research is consumed. A state-of-the-art paper from Meta or TikTok is a recipe calibrated to behavioral data most companies will never collect, so copying it imports the technique but not the moat. The real differentiator was never the algorithm; it was whose data the algorithm was tuned on.
03The Core Invention

Large Event Models, Not an LLM

Sequen's starting point was a model that generalizes to human behavior the way a language model generalizes to text, built deliberately without any large language model or generative AI.

it's not a large language model. We actually don't use large language models anywhere. That's not our business whatsoever.

— Zoe Weil, The TWIML AI Podcast
Key Insight
Naming it a large event model is a deliberate bet that the LLM playbook, scale a sequence model until it generalizes, carries over to behavior. But she strips out the LLM itself. The claim is that the portable part is the architecture lesson, self-attention over a sequence, even when the tokens are hovers rather than words, and that the generative layer everyone else is racing on is simply irrelevant to ranking.
04The Signal

Look Between the Clicks

Most teams optimize the rare explicit signals, a click or a thumbs-up; Sequen relies overwhelmingly on implicit micro-signals like dwell time, scroll speed, and how many times you go back and forth before the cart.

I'm talking about hovers, scrolls. I'm actually talking even deeper than that. The speed of a scroll, right? The dwell time on a hover.

— Zoe Weil, The TWIML AI Podcast
Key Insight
Clicks and thumbs-ups are explicit feedback: rare, and biased toward the small slice of people who already converted. The signals she wants are produced by everyone, including the roughly 99% who leave without buying. The unlock is not more data about a few buyers; it is any data at all about the many who never register in a conversion log, which is exactly the population a transaction-only view is blind to.
05Privacy by Architecture

You Are a Session, Not a Name

To Sequen a user is a collection of events in a live session, not a stored identity, so there is no name, location, or cross-site profile to hold onto in the first place.

When I think about a user, think about it as a collection of events in a session. I don't care if you're John Smith or Adam, whatever.

— Zoe Weil, The TWIML AI Podcast
Key Insight
Her answer to the surveillance-capitalism charge is architectural, not just a policy promise. She describes the model's unit as a live session rather than a person, and says Sequen does not store people's information or care about a user's name or location. If the system genuinely needs only in-session events, then the durable profile that would make it surveillance is also the thing it was built to work without, so the incentive and the ethics happen to line up rather than fight.
06The Latency Moat

Under 50 Milliseconds, in Rust

Cleaning raw events, passing them through hundreds of models, and serving a re-ranked page in under 50 milliseconds at thousands of requests per second is what forced the core backend into Rust rather than Python.

the core of our backend is all in Rust. So no, I can definitively say we could not have done it in Python at all.

— Zoe Weil, The TWIML AI Podcast
Key Insight
In real-time ranking the hard problem is not the model, it is doing clean, leakage-free feature transforms and inference inside the few milliseconds a page takes to render, while reacting to signals that are still arriving. That budget, not the algorithm, decides who can play. It is also why she credits luck and timing: open-source Rust support for ML matured right as they needed it.
07The Durable Moat

Residual Knowledge Beats the Algorithm

Making these models work takes an innate sense of which features drive value that is not written in any research paper, so Sequen hires it out of big labs and codifies it into a reusable knowledge base.

requires you to have an innate understanding of what features drive value and that is not written in any research paper.

— Zoe Weil, The TWIML AI Podcast
Key Insight
Her claim inverts the usual faith in open research: published papers are a lagging, sanitized signal, because the strongest methods sit behind months of legal review and ship in watered-down form. On top of the proprietary models and recursive stack she calls the company's core IP, residual knowledge adds a moat a competitor cannot simply download: accumulated, codified intuition about which features matter, the part that never reaches a PDF.
08The Compounding Engine

Why It's Called Recursive

Early on Sequen retrained ranking models from scratch, sometimes hourly; the recursive stack instead learns continuously with small updates layered on prior ones, compounding gains at a fraction of the compute.

you don't have to rebuild the entire ranking stack every time

— Zoe Weil, The TWIML AI Podcast
Key Insight
Recursive is doing real work here, not branding. She contrasts retraining the ranking stack from scratch every cycle with layering small updates on top of prior ones, so improvements can build on each other instead of resetting each time. That compounding at a fraction of the compute is the mechanism she credits for moving fast: the same reason interest compounds, applied to model quality instead of money.