We’ve just launched our agentic capabilities to make your AI workflows even more powerful. Try it out for a limited time →
Q: Can you introduce yourself to our audience?
Muskan: My name is Muskan Kukreja. I'm the head of ML and AI at Checkr, a leading tech-based background check company. Our mission is to build trust between two parties.
Q: How do you guys at Checkr use ML and AI, both traditional ML and generative stuff?
Muskan: Typically I like to basically say why using ML and AI is not really the first thing that we think about. The whole thing started with this industry being pretty boring, done by two or three providers in the world. Frankly there was not much need for it - employment was slow, people used to get hired in 10-14 days.
It really started with this concept called gig economy where there has been a need to get somebody employed the same day, in thousands of numbers. Not just in the tech economy, but driving Uber, nursing, hospitality, construction. The world has moved from sticking to a job for 40 years to changing jobs every single day.
We had to step back and rethink the current solution. That's the main reason Checkr was born - trying to redefine how we establish trust between two parties with accuracy, thoroughness, and high quality responses in the fastest amount of time possible.
All of our use cases in AI, gen AI or ML are based on whether a particular solution can revolutionize this industry by applying ML and AI. Is it going to increase our accuracy? Is it going to improve our turnaround time? Is it going to decrease our cost? Cost is actually a growth play - imagine if I can offer a background check for $1 in 5 minutes, you would probably do it for your landscaper or housekeeper.
Q: What's the kind of scale that Checkr is operating at? Who are some logo customers that people will know?
Muskan: We operate globally, but the US is our first market. We serve all segments from SMB to Enterprise. Some name customers include Uber, Instacart, DoorDash - all the gig economy companies are our customers. But we've built our solution for pretty much everybody, so we have a very good amount of representation from every industry, thousands of customers, including consumers too.
Q: What does the head of ML and AI actually do at Checkr?
Muskan: Thanks to the leaders of the company, I was interviewing for the head of ML position, but we want to be more inclusive to ML and AI here. We started with a couple of people who were pushing boundaries with ML, and I saw a lot of promise. We thought about establishing and growing that arm of the company where we can create a diverse skill set in machine learning, data science, and AI, and also leverage some of our current software engineers and train them to be more AI-driven.
We have two functions in the company. One is to uplift everybody to be more AI-driven - teaching them, helping with their current jobs, employee productivity stuff. The second is to support different pillars across the company, focusing on two things:
First, can we use ML in different internal processes that aren't customer-facing to boost our accuracy, optimize, and reduce our cost of searches? These are backend processes where we're doing data processing and trying to make that process more efficient.
Second, we focus on the customer. Customers look at complicated background check reports and want an AI solution to explain them, justify why something shows up in their report, help with crafting stories, summarize legal complicated text, and provide recommendations like which packages to order for a given job description. This is the ML that our customers or candidates see, feel, and touch - the external facing AI.
Q: How did you guys get Checkr to be so sophisticated with generative AI adoption?
Muskan: I have to give credit to the leaders of the company - my VP of Engineering, Christa, Co-founder Daniel, and other leadership people. It has to be a top decision because if we didn't have tools like Credal in our company, even if we wanted to do something, we couldn't.
It starts from the top - acknowledging that this technology is not a fad but has legs, doing due diligence, selecting the right tools and partners to work with us, and provisioning that system inside the company. It's not cheap to do this for the whole company.
Having the right mindset is the first thing, and then creating an AI enablement team - a set of people who will be your power users who can uplift everybody and get into the use cases. Having that right culture and flexibility is important.
We did a hack week where the whole company for one week was not doing anything except AI. Even if I wanted that, it couldn't have happened without company buy-in. We were responsible for teaching people how to be at the forefront in using all these tools.
Q: What were the learnings from the hackathon that helped keep confidence that this was a good investment?
Muskan: When I came in, this was Daniel and Christa's idea. I was not actually sure it was a good decision - maybe we should just use two tracks, AI track or non-AI track. But fixing on the learning part got everybody excited.
When you go to a bar in San Francisco, people are talking about ChatGPT, and even the smartest AI engineers - when conversation of AI happens, people who don't have context feel out of the room. There's a big FOMO.
I thought in a week I could uplift anyone to be at least 30% more effective, and now they can have a meaningful conversation about AI. I just needed the right curriculum and right people. That's where the conviction came - let's force people to participate with learning first.
We spent almost 20% of time teaching people tools like Credal, how to create co-pilots, when to use them. About 25-30 projects came up that I wouldn't have thought about, especially from people who aren't even in tech. People from marketing and different disciplines who had never written a single piece of code were now doing co-pilots and uplifting everybody.
Q: What are some of the most exciting use cases you've seen?
Muskan: When I came here, I mapped my entire team into four parts based on business units, which has some resonance with how we think about AI.
One part maps to internal needs. Every company, especially in engineering, deals with technical debt. Rearchitecting current systems and simplifying that is a process every company goes through. We have initiatives on how we can simplify tech debt using AI.
Generating code is one of the biggest use cases we're seeing. Amazon listed some crazy numbers of how AI saved 4,500 years of development time. Google said 25% of their code is written by AI. Generating code and AI helping with development is a critical component.
Running a background check involves searching for people's identities in multiple places, which requires tons of operational work, especially the manual QA process when something isn't working. The biggest use case we've had success with is replacing that manual burden in terms of cost, time, and accuracy, because humans can be inconsistent - they get sleepy, they're expensive in the US.
Q: How does the manual QA process work today?
Muskan: One of the biggest use cases is our verification arm. When we do verification, we end up with thousands of documents - degrees, PDFs of employment W2s, driver's licenses. We started with traditional OCR, but one bottleneck is that the tech has to support a known document type, or we have to train by artificially generating data.
One of the first AI use cases was looking at a W2 employer and matching it with what we found, putting the title and matching it. The surprising thing was handling cases like when someone says they're a product leader at a company but their title is program manager - is that a bad match? This requires human judgment, but AI is pretty good at mapping different job titles and figuring out if a document belongs to someone, if they really worked for the claimed time period.
Document verification and understanding was one of the biggest use cases. The backbone of why Checkr is an ML and AI company is that we're dealing with text data. Every company has use cases for code generation, documents, marketing campaigns. But if your company deals with text data, you cannot survive in this market without being an ML company - especially when this text is not natural language but complicated legal data.
Q: How do you manage hallucination risks and other concerns for these high-stakes customer-facing use cases?
Muskan: This is a very sensitive topic. The biggest success any company would see, including Checkr, is to not start with an AI use case for customer-facing applications, and not start with text generation that could impact people's lives.
We want to start with use cases where we're learning about the potential of AI and where mistakes aren't costly. Any process which was human-driven that we're replacing or augmenting with AI still has humans in the loop.
For example, with title matching, we're replacing human matching with AI matching that's thoroughly reviewed. We can review the combinations of titles that are matched, pre-fetch and pre-cache them, and when new titles come in, we can employ that approach. When we're not confident, we can have humans in the loop.
A good AI use case is extremely low risk, extremely efficient, saves money, and is easy to implement. That's why we've been doing AI internally first, to understand what we know and don't know.
We've established a process with a panel in the company - a checklist for any ML or AI feature that wants to roll out. It has to go through certain processes to make sure we're using data properly, there's no PII involved, and we're following security, legal, and ethics requirements.
We started with boring use cases: Can we help detect anomalies in our data pipelines? Can we help with document verification with humans in the loop? Can we use AI to generate recommendations like package options? The human is still in control.
Q: It seems like focusing on boring use cases is actually a smart approach?
Muskan: Whenever you're in an industry that gets a lot of hype and buzziness, very often the correct approach is to focus on the most boring possible thing. If you take that mindset, you filter out all the noise of the super buzzy demos and viral Twitter content that doesn't work, and you can focus on something simple that's actually concretely valuable for your users and employees.
You tackle these boring things and gradually build up from there into more impressive things. Even with traditional machine learning, there was all this hype about predicting everything with impressive models, but what turned out to be valuable was spending time cleaning up data and using it for something very narrow, specific, and useful.
In machine learning, when we're in college, we're given a dataset and focus our energy on models. But in industry, we fix models based on what's supported in the platform, and the real benefit comes from the data. There's a tradeoff between data and models.
That's how we started with AI. Being in a novel department that doesn't ship anything for 2 years isn't ideal. We take long-term strategic bets, but I always look for the lowest hanging fruit that we can do in less than a quarter while continuously keeping milestones for long-term projects.
Q: How do you create a gen AI/ML team that can succeed? What skillsets do you look for?
Muskan: This is very interesting, and every company is slightly different. The talent is moving in a different direction. My philosophy has always been to understand what a company wants from an AI team.
When a business problem comes, it doesn't have a face of what technology it is - the problem is a problem. When you convert a business problem to a technical problem, then you realize it's an ML problem or an AI problem.
I create a diverse team of skill sets - what we call a "T-shaped team." In a small company, everyone has skills across different areas. In a big company, there might be seven different titles - data engineering, analytics engineering, data science, data analyst, applied sciences, research scientist, machine learning engineer, AI engineer, software engineer. If you have $3 trillion in revenue, you can hire a team like that, but how do you do that as a startup?
You create a diverse team of skill sets - people who are modelers, people who love building evaluations or feature stores or infrastructure to deploy models. We look for people with great curiosity, excellent soft skills, and we build a diverse team across different use cases. When a business problem comes, I'm confident I have the team with the right skill set, whether it's statistics, experimentation, ML, or AI.
Q: How does Credal fit into your AI strategy?
Muskan: I have to give credit to the leaders of the company. If we didn't have tools like Credal, even if we wanted to do something, we couldn't. It starts from the top - acknowledging that this technology has legs, doing due diligence, selecting the right tools and partners, and provisioning that system inside the company.
The creation of co-pilots has been significant - we have so many co-pilots running. Earlier, people used to ask questions about a project. Now we create richer documentation, create a co-pilot, deploy it to Slack, and enable it automatically. When people ask what an algorithm is about or where we get data from, the Credal Slack bot thinks, fetches the documents, and gives the answer without us having to look at those responses.
People have created co-pilots on hiring in Canada and other topics. When you have institutional knowledge, you create a document, create a co-pilot, put it on Slack or different places, and you've democratized your information and saved tons of money.
One reason for Snowflake's success was that it was cloud-agnostic. Similarly, Credal is like "you like Anthropic, you like OpenAI, you like Gemini, you're not sure, you want flexibility to switch in and out, and you don't want your users to care about and procure five different things." Having that one layer to manage security, enterprise RAG, the ability to switch models, and fast access to the API with no breakage is valuable.
Q: What are you most excited about building with AI over the next 3-12 months?
Muskan: I've created four workloads and four parts, and I'm really looking forward to simplifying the data architecture. Companies like Checkr rely on thousands of signals, sources, APIs, and architecture to serve a customer request.
We have scrapers and parsers running all day scraping websites, all AI-powered. One of the biggest use cases I'm looking forward to is self-healing AI-generated code. Imagine you write a parser using Beautiful Soup, and then the website changes its layout - your parser breaks. We've prototyped code that can self-heal based on layout changes.
Support is another area - if you don't find an AI use case in your company, you haven't looked at support. Customer support tickets contain text, and you want to quantify emotion, sentiment - all the typical textbook use cases of AI.
I want to take maximum advantage of established use cases to ensure all our basics are covered, and then tackle actual user delight. We're currently using ML algorithms for recommendations and fraud detection, but I want to convert those to AI.
With multimodal AI, I'm rethinking whether we even need OCR or if we can have cheaper AI just reading documents and verifying information, with humans in the loop to guide the process.
Q: What advice would you give to someone building an AI platform team?
Muskan: In any career, especially ML and AI, if you find good business problems to solve, you're set. The question is where do you find those business problems? I encourage engineers and scientists to understand your company goals, priorities, roadmaps, and strategy so you can identify the good problems. You're closer to the data and technology, but you need to understand the business goals and objectives. If you get the right problem in hand, you have the toolkit to create wonders.
Credal gives you everything you need to supercharge your business using generative AI, securely.
