Emily Glassberg Sands joined Stripe in 2021 to lead data science after Stripe CTO David Singleton contacted her.
“I joined Stripe, uh, back in 2021, originally actually to lead data science, and David Singleton, Stripe's CTO, reached out.”
Person
Emily Glassberg Sands
1 episodes · 24 quoted claims
Three Stripe engineers built an internal beta of LLM Explorer in three weeks.
“three engineers who hacked together in three weeks an internal beta for an LLM explorer”
Stripe built security features such as stripping personally identifiable information from prompts and restoring it afterward into LLM Explorer from the start.
“We knew we needed to have certain security features, stripping PII and rehydrating, et cetera, straight from the start.”
About a third of Stripe employees used LLM Explorer within days of launch, and the tool works as a front end over multiple models.
“Within days, a third of Stripe's were using it. And you can think of LLM Explorer basically as a front end that supports multiple models in the back end.”
LLM Explorer launched with GPT-3.5 and GPT-4 and now serves more than six models.
“So we started just with GPT 3.5 and GPT four, but today we serve over a half dozen models through the tool.”
After Stripe added a presets feature to LLM Explorer, about 300 reusable prompt patterns appeared almost overnight.
“basically, overnight, we had like 300 of these reusable LLM interaction patterns, and they ran the gamut.”
Thousands of Stripe employees use an LLM Explorer preset called the Stripe style guide, which rewrites content in Stripe's tone.
“thousands of stripes still today use the stripe style guide, which basically, you know.”
LLM Explorer has almost 3,000 weekly active users, close to half of Stripe's workforce.
“the weekly active user count of this LLM Explorer is still at almost 3,000, which is just shy of half the company using it every single week.”
Stripe's internal accelerator program funds small one-pizza teams for about six months with defined charters and milestones.
“we'd like to build a one pizza team and we we wanna fund it for six months, so relatively durable.”
Stripe Radar was one of Stripe's first non-payments products and uses machine learning to identify and block fraudulent transactions.
“It was one of our first non-payments products, it's an ML powered product that helps identify and block fraudulent transactions.”
Sigma is Stripe's SQL-based reporting product, which lets businesses analyze their own Stripe data.
“So Sigma is our SQL based reporting product. Um, it lets businesses analyze and get insights directly from their Stripe data.”
About sixty applications have been built on Stripe's internal LLM API, both internal and external.
“There are sixty applications built on that now, a bunch internal, but also several external, and I'm happy to talk about a couple of them.”
Stripe bills the most expensive LLM applications to the teams that own them rather than absorbing the cost centrally.
“we found that for, um, the most expensive applications, you know, we do bill them to the local teams.”
Stripe built its experimentation platform internally because charge-level experiments need very high latency and reliability standards.
“Our experimentation platform is one that we've built internally. We run a lot of charge level experiments, and latency and reliability requirements for charge level experiments are very, very high.”
Sands speculates that a foundation model trained on financial data could beat traditional payments optimization models by roughly 100 to 200 basis points.
“it doesn't feel crazy to think that a good foundation model could outperform more traditional approaches by, I don't know, a hundred bips, 200 bips.”
Stripe's optimized checkout suite bundles front-end optimizations such as showing the most relevant payment methods first.
“it's this bundle of front end payments optimizations, and it's a lot of little things, honestly, like dynamically presenting payment methods in the order that are most relevant for the customer.”
Stripe uses machine learning to optimize authorization requests and retry routing to recover a large share of falsely declined payments.
“we use ML to optimize authorization requests for issuers, basically identifying the optimized retry messaging and routing combinations. Uh, to recover a big chunk of false declines”
Stripe's smart dunning in its billing product picks the best day and time to retry declined recurring payments, such as those that fail for insufficient funds.
“You basically identify the optimal day and time to retry a payment, um, for transactions that are declined, for example, due to insufficient funds.”
Stripe Radar evaluates about a thousand characteristics of each transaction in under a hundred milliseconds.
“Stripe Radar, which you mentioned, you know, considers a thousand characteristics of a transaction and figures out in less than a hundred milliseconds.”
As a college student, Sands ran an audit study sending four unseen scripts to hundreds of theaters while varying the author's pen name.
“some excellent playwrights donated four never-before-seen scripts. And I sent them out to hundreds of theaters and asked them, um, whether they wanted to put it on stage, why or why not?”
Sands's audit study found that the same script was less likely to be produced when its author appeared to be a woman.
“basically, I found that when purportedly written by a woman, the exact same script was less likely to be produced.”
Sands spent about eight years at Coursera, rising from an individual contributor to leading the end-to-end data team.
“I was at Coursera for about eight years. I grew from an IC to leading the end-to-end data team.”
Sands says AI startups face substantial compute costs from the start, which pushes them to build monetization sooner than earlier software startups.
“unlike a bunch of the past generations of software startups, we're seeing AI startups have substantial compute costs right out of the gate.”
Sands says more than half of Forbes's list of top fifty AI companies from the previous year used Stripe.
“you know, we're looking at the list of top fifty AI companies put out by Forbes last year, and notice over half were using Stripe.”