Research Engineer, Content Understanding
At a glance
Mid-level AI Research role at Exa. San Francisco · full-time · $180,000–$350,000 base.
aimlroles summary, based on the employer's posting.
What you'll do
- Improve page parsing by separating meaningful content from layout elements, then demonstrate that the results got better
- Develop models to assess page quality, credibility and misinformation, including whether content serves readers or search crawlers
- Create web-scale classification and extraction that stays inexpensive enough to apply broadly
- Build supervision for tasks without established labels and test whether models can learn them
- Trace poor search results to the page-level predictions behind them and correct the underlying issue
What you bring
- Graduate-level machine learning work plus at least two years of relevant experience, or an exceptionally strong undergraduate background
- You can implement a transformer from the ground up in PyTorch and train models designed for low-cost use
- You enjoy building large datasets and shaping the supervision that drives model performance
Who this fits
This role suits a research engineer comfortable defining ground truth for open-ended problems and working on large-scale web content understanding. It is an onsite position in San Francisco.
From the employer
Exa is an applied AI lab building a search engine unlike the world has ever seen. We build massive-scale infra to crawl the entire web, train state-of-the-art embedding models to process it, and design super high performant vector databases to retrieve over it. We now power search for Cursor, Cognition, HubSpot, and over 400,000 developers and have raised $350m from Lightspeed, Benchmark, and a16z.
Our ultimate goal is to build perfect search over all the world's information, far beyond Google. If you want to build massive-scale ML systems that will define the way the new AI world consumes information, this is the place for you.
As a backend engineer, you'd play a critical role in our search architecture. We're pretty flexible on what projects people work on based on their skills and interests.
Search quality is bounded by what we understand about a page. Before anything can be retrieved, something has to work out what the page actually says. That means parsing it into the parts that are content and the parts that are furniture, classifying what kind of page it is and what it is about, telling whether the page is usable at all, extracting when it was published, judging how good it is and whether it can be trusted, and working out whether it says anything that a page we already have does not. All of this has to work on every page on the web, in every language, in every shape the web comes in.
Some of this is classic document understanding. Some of it is much more open. Credibility and misinformation, AI-generated and machine-spun content, and pages written to be found rather than read are all unsolved, and search results are only as trustworthy as our answers to them.
We are looking for a research engineer to work on this. There is a lot of room to do it well.
Desired Experience
Graduate-level ML experience (Master’s or PhD with at least 2 years of relevant experience), or an exceptionally strong undergrad
You can build a transformer from scratch in PyTorch, and you have trained models that then had to be cheap enough to run everywhere
You like building large-scale datasets and living in the data. Most of the wins here are in the supervision rather than the architecture
You are comfortable with problems where the ground truth does not exist yet and defining it is part of the job
You care about the problem of finding high quality knowledge and recognize how important this is for the world
Example Projects
Make parsing work on the pages where it currently does not, and prove the improvement rather than assert it
Teach a model to judge page quality, and get everyone to agree on what quality means well enough to supervise it
Work on credibility and misinformation as a modelling problem: what a page claims, whether it is a reliable source of it, and whether it was written for a reader or for a crawler
Decide whether two documents are semantically the same or genuinely different, so we can deduplicate the web without collapsing pages that a user would want to see separately
Build classification and extraction that is accurate at web scale and cheap enough to run on all of it
Design the supervision for something nobody has labels for, and find out whether it is learnable at all
Trace a bad search result back to the page-level prediction that caused it, and fix it at the source
Exa is an equal opportunity employer. All qualified applicants will receive consideration for employment without regard to race, creed, color, religion, sex, sexual orientation, gender identity or expression, national origin, disability, age, veteran status, marital status, pregnancy or related conditions, criminal histories consistent with applicable law, or any other basis protected by applicable law.