To build AI you need algorithms, compute, and data. While we have collaborative tools for tasks like writing and debugging code, the machine learning world has no standard tooling for labeling data, storing it, debugging models, or continually improving their accuracy. Labelbox solves the bottleneck in ML by providing a complete solution with fast labeling tools, human workforce, data management, a powerful API, and automation features.
Each team is asked to select, explain, and rank their top 8 values in order of importance.
Customer Comes First
Labelbox helps any team of data scientists who wants to get a computer vision model into production.
Before starting Labelbox, Manu (our CEO) worked at Planet Labs where he led a team that spent millions of dollars trying to make sense of training data. We interviewed a dozen companies who were all struggling to build machine learning products and services, and realized how widespread these challenges were.
Today, our customers are as wide ranging as the applications of AI. Accurately and quickly labeling source data is paramount, regardless of the use case. One of our customers is building an AI model from satellite images of trees in order to prevent fires and save lives. Another is ingesting tens of thousands of radiology exams to produce cancer-spotting algorithms. Labelbox is used to further innovation in agriculture, medicine, insurance, and even dental fields.
We believe artificial intelligence has the power to transform every aspect of our lives, from healthcare to agriculture. Learning about new use cases for Labelbox and new ways of expanding our impact is never-ending. We’re excited that our product is accelerating the development of AI, which in turn accelerates our customers’ ability to quickly and accurately detect cancer, identify when disease hits their farms, and build safe self-driving cars. Labelbox is a catalyst and we pride ourselves on saving companies from having to create their own expensive and incomplete homegrown tools. Instead, they can rely on our training data platform that acts as a central hub for humans to interface with AI. When humans have better ways to input and manage data, machines have better ways to learn.
While it’s great if you’ve personally felt the problems we’re solving for, we don’t need you to be intimately familiar with machine learning. It’s more important that you are excited about the future of machine learning and can get behind our views on craftsmanship.
High Quality Code Base
Craftsmanship to us means communicating through your product.
One of our core values is craftsmanship. The way you see the world shapes how you build your product, and your users will then see the world through your eyes when they use your product. This is incredibly powerful. Our perspectives on the world and even our emotions can traverse lines of code and be communicated through the things we make for others. We are engineers and product managers, but above all, we are craftsmen and women.
Our engineering team focuses on shipping reliable features with the aim to continue building and growing our customers’ trust in Labelbox. We consider ourselves to be partners with our customers. At our current stage, we still do one-off feature builds for individual customers in order to help them leverage our product as much as possible. We will likely grow out of this as we scale, but for now, this has proven to be the best for both our customers’ success and our own understanding of how people use Labelbox.
We pride ourselves on writing high-quality code that is both easy to read and maintain. We operate using agile processes and two-week sprints. Each engineer owns a product for six weeks and we have a backlog of stories that people can pull from. Part of what contributes to our overall code quality is iteration. We do code reviews, deploy 4-5 times per week, and try to get things out quickly in order to get feedback that informs us on what to improve. We use LaunchDarkly for feature flagging which helps us roll our new features quickly and safely. Perfection never blocks us, but we do aim for it.
Our ownership mentality is best demonstrated by how we do OKRs.
Every engineer at Labelbox owns a single OKR. They are responsible for everything related to their OKR – design, customer success, development – and after six weeks, we come together to switch up OKRs for the next six weeks.
Everyone is involved in our product roadmap. Features tend to come together from multiple people on the team, each person contributing their insight, input, and ideas. We are currently ~15 engineers which means that everyone has the space to reach and stretch.
Our image segmentation feature is a great example of how features come together: the project was divided into smaller parts and each engineer was able to own the development process. That process included talking to customers, making case files, designing, and then implementing this feature. The end result of this effort is a product that allows customers to label every pixel of an image. Think Photoshop for the browser.
We believe that having ownership means having agency over your schedule, too. Labelbox is 50% remote and we have engineers on multiple continents. Everyone is fairly autonomous. You can roll in at 8am or 10am in your local timezone, go to the gym in the middle of the day, and structure your schedule in whatever way makes you productive and happy.
We spend a lot of time on design.
Since the very beginning of Labelbox, we’ve had a design-oriented culture. When we brought on our Head of Design David from Dropbox, he helped bring a more methodical approach to how we do design at Labelbox. (David said they’d sometimes do 100 iterations before moving forward with a design at Dropbox.) Like most early stage startups, the question of “how can we go faster?” comes up often. To us, it is much easier to make 10% better decisions than making engineering 10% faster. We invest a lot of time in designing things the right way before and will do a lot of iterations before we write a single line of code. (Maybe not a 100, but still a lot!)
We recently did a design sprint where five team members set aside four entire days to just focus on how to improve labeling speed. The team's efforts are now powering a new strategic direction and we are optimistic about doing more design sprints.
How are we going to win the market? By having the best product.
In a lot of ways, we are creating an entirely new category. Today, machine learning teams spend ~80% of their time creating and managing training data. We’re offering best-in-class tooling, collaboration, and dedicated labeling services as the first training data solution for machine learning, and even though we’re the first to do it, we’re going to do it really well.
If you consider where the world is moving and what the competitive landscape looks like in artificial intelligence, there are many ways to “win.” You can solve the hardest engineering problem, or you can build software to help people collaborate. Some companies are working to have the best label services, but our plan is to build the best software product that will be the GitHub for training data. Rather than having every company create their own expensive and incomplete homegrown tools, we’ve created a training data platform that acts as a central hub for humans to interface with AI. When humans have better ways to input and manage data, machines have better ways to learn.
Friends Outside of Work
We started the company by hiring through our networks, so there are lots of friends here!
Five of our earliest hires loved working at Labelbox so much that they brought on their best friends, so we literally have a lot of best friends at the company. Of course, as the company grows, we don’t know if this will continue to be true, but it’s definitely true now and we hope we’ll continue to be the type of company where people want to recruit their friends to.
We generally enjoy each other’s company and like organizing events for us to do together. We pass around The Burrito Ring every month, which is essentially us taking turns to organize a company event like camping at Yosemite or throwing a boat party (we have three DJs at Labelbox, a complete coincidence, we swear). The only rule for the event is “Be Responsible,” which hasn’t been abused yet, but we’ll see. 😂 Every other Friday, we have company demos (where everyone shares what they’ve worked on), and then after, we have a BBQ with music, snacks, wine, and hang out on our patio. Every quarter, we have a Magic the Gathering night where we all get starter decks and play together, and every year, we have a Christmas party. Really, we like finding reasons to hang out because we really do like each other.
To be clear, you do not need to be friends with anyone here in order to join us! While we do two culture interviews, they’re not about shared interests or who you know. We care much more about your communication and conflict resolution style. That said, there are no promises that you won’t end up making lots of new friends once you do join.
Labelbox has an automated pipeline that allows us to deploy multiple times per week (or per day).
We use Codefresh so our tests and deployments take less than 10 minutes. One reason this is such an advantage is because continuous delivery allows us to test what works. To us, continuous delivery is saying, “I don’t know all the answers, and I need to capture it through experiments and iteration,” and asking, “Does this work?” over and over again. We’re also confident every time we roll out changes or new features because we use LaunchDarkly for feature flagging.
Customer Comes First
Bonded by Love for Product
High Quality Code Base
Friends Outside of Work
12 Full-Stack Engineers
2 Product Managers
On the frontend, we use TypeScript, React, Redux, GraphQL, Apollo. On the backend, we use Node.js, Go, TypeScript, MySQL, Elasticsearch, Redis. Everything is containerized in Docker and running on Kubernetes.
1. Hello/intro call.
2. 30-min coding challenge, asynchronous.
3. 60-min pair-programming problem done over google hangout.
4. Onsite, 4-hours long, which includes 2 culture interviews, 1 architectural interview, and 1 final recap/final meeting.
5. We’ll meet internally, do reference checks, and then send a written offer.