Karim Atiyeh and Nik Koblov (Ramp) Fireside Chat

February’s Chat8VC marked our first NYC gathering, and the Ramp team was instrumental, hosting us in their beautiful space, demonstrating their latest product, and headlining the fireside chat. As a reminder, Chat8VC is a monthly series, most often hosted at our SF office, where we bring together founders and early-phase operators excited about the foundation model ecosystem. We feature highly technical, implementation-oriented conversations, and give founders and builders the opportunity to showcase their work through lightning demos. This can be a side project, related to a company they’re starting, or a high-leverage use case in a company operating at scale.

In conversation with 8VC Partner Bhaskar “BG” Ghosh, Ramp’s Co-founder & CTO Karim Atiyeh, and Head of Engineering Nik Koblov shed light on the early days of Ramp, how the same mission has guided them since well before AI was a major factor, models and infrastructure, building a magnetic talent culture, and more.

‍If any of these themes excites you and you’d like to chat more with our team about them or attend future Chat8VC events, please reach out to Vivek Gopalan at vivek@8vc.com and Bela Becerra at bela@8vc.com!

We are so fortunate to engage in a fireside chat with Karim Atiyeh and Nik Koblov of Ramp. Let’s kick off with brief introductions and then we can launch into rapid-fire questions.

Karim: I'm Karim Atiyeh, one of the founders of Ramp. Prior to Ramp, my co-founder Eric Glyman and I started Paribus in 2014 – funny enough, it was the first version of what we now refer to as an “agent.” Every time I hear “agent,” I have to reconfirm what exactly it means. At that time, it was much harder to build. Paribus connected to your inbox, identified receipts, the items in those receipts, then scraped prices, recognized price drops, and reached out to stores on your behalf via email, with the ultimate goal of getting you money back. 

It was challenging and took more effort, seeing as we didn’t have access to today’s tools. The experience was incredible, though, because it was our introduction to entrepreneurship. We sold Paribus to Capital One – we learned a lot about how credit cards work, what it's like to operate in financial services as well as the associated regulatory challenges, but most importantly, we recognized the massive size of the opportunity.

Shortly after that we kicked off the Ramp journey, which is one that we’ll likely be on for many years to come! The initial idea for Ramp was inspired by Paribus: Could we build Paribus for businesses, and help businesses save money automatically? We are a spend management solution that helps companies manage purchases, reimbursements, procurement, and travel spend. We empower business leaders to make better, data-driven decisions in order to save time and money. 

Corporate cards are a tool, but there's a lot more that goes into making decisions around what, why, and when you transact. 

Nik: My name is Nik Koblov and I'm the Head of Engineering at Ramp. I met Karim back in 2019. I had just wrapped up serving as the Head of Bank and New Markets Engineering at Affirm. I developed a passion for consumer fintech, having built most of the rails, BNPL, and lending engine as well as bank account features. We expanded to NY, and I coincidentally bumped into Karim, who had launched Ramp eight months prior. We were catching up and he whipped out the Ramp card and walked me through the user experience. I was impressed by their velocity of execution and ambitious vision. 

To be completely candid, though, the corporate card idea was personally a detractor. I didn’t particularly want to build corporate cards, seeing as they’re historically terrible products. However, I was blown away by the caliber of talent Karim had recruited, as well as how quickly they were shipping new features. I asked Eric “why build Ramp?”, and he noted that he wanted to help companies run leaner. The combination of a clear mission, talent magnetism, and Karim and Eric’s humility inspired me to join.  

While Karim and Eric understood the opportunity quite well, they demonstrated very low ego around the long-term product vision, always prioritizing a data and customer-driven approach to inform the roadmap. Rapid feedback loops have always been at the core of Ramp’s DNA – discovery by getting products quickly into customer hands, realizing what they need most and how they're going to use it to grow their business. I've never seen a company iterate faster than Ramp. 

Prior to Affim, I learned about fintech at Goldman, so I'm very much the guy who brings the “fin” to this tech company. I dabbled in startups before Goldman and studied physics. I came from a family of scientists and was a bit of a rebel as I was the first engineer in the family. 

VCs are coached to ask this facile question of “why now?”. And I scratched my head and I couldn't figure out why on earth would you start a credit card and expense management company. What was the impetus for Ramp – what gave you the confidence you could do something different? 

Karim: Timing is critical. It’s tough to predict, but sometimes you look in retrospect and it’s so clear the timing was right. Given the 2008 financial crisis, a lot of regulation was instituted to ensure there weren't as many banks that were too big to fail. In certain cases, the regulators tried to make it so that smaller banks had some competitive advantage. The government allowed smaller banks to keep more revenue in the form of interchange in certain consumer use cases.

For a few years, there was a wave of infra startups that aimed to help old-school small banks who were keen to capitalize on this competitive advantage. Banking-as-a-service type companies started to emerge – these are players that helped banks incorporate tech and collaborate with startups. We are part of the second wave, as there was already a significant amount of infrastructure to connect to all these banks, without being dependent on any one of them. We can innovate and actually focus on the tech side, without having to deal with the slow pace that comes with regulation. 

It was a good time to start this type of company. And capital was also very cheap. 

Before we segue to AI, was picking small business credit cards and optimizing for spend management the heart of the initial vision? 

Karim: The initial vision was oriented around the best way to save companies time and money. For consumers, we thought we could convince individuals one by one to give us access to their mailbox – it’s quite hard to do. You have a lot of information in an inbox around spend and the opportunities to save. It would be difficult to convince businesses to do this first. But, you also have a lot of information on what businesses are spending on, from credit card data. Cards are quite antiquated – they haven't evolved in 20 to 30 years. Most credit card companies spend very little on innovation. Most of the banks and companies innovating in this ecosystem only know how to compete on points, rewards and marketing.

We set out to build the card, and compete by offering a superior product as well as a clear focus to empower businesses to save time and money.

Nik, when you first started at Ramp, what were the key areas where the team was already using AI or ML in general? If you look back four years ago, what was the core AI enablement vision?

Nik: It certainly wasn’t at the heart of our DNA, as savings was the key focus. The initial goal was to give cards to businesses to understand what they're spending on, and then produce insights around the trends in your spend, flagging outliers as well as opportunities to save. 

We also invested in partnerships early on so we could provide meaningful discounts. Buying a SaaS subscription through Ramp gets you a 5% discount, for example. Generating tangible savings from day one was always our North Star and core component of the vision – we always need to drive value. 

Were you using classical machine learning models at that time?

Karim: Not as much at that time. It was more about heuristics, though we built a receipt-matching product quite early, where we were using OCR, trying to figure out what was and wasn’t a receipt. 

There was clearly a fair amount of unstructured data and text already flowing through the platform. And this predated the whole Generative AI wave, so what were the critical components then? 

Nik: Our first trained ML model was in the risk space. We trained a model that determines the optimal credit limit, which is based on the probability of a delinquency, and conditions around the exposure that Ramp is willing to take for that account. We needed to get a fair amount of delinquencies to train - there was always a Catch-22 with fintech. At Affirm we had a budget allocated for random underwriting – the scale was tens of millions of consumers. At Ramp, the dataset that became statistically significant first was the number of transactions. We got to tens of millions of transactions quickly and then trained a ML model for fraud detection. 

We handle transaction-level underwriting and determine the most optimal limit. The goal is not to prevent losses, but to generate the right risk-adjusted growth.

Karim: We actually started using AI in our growth team very early on. We built an incredible top-of-funnel engine to automatically reach out to prospects. We first captured the data from people visiting our site, enriched it,  then reached out to prospects with very personalized messaging. 

Let’s briefly touch on document processing as well as tabular and non-tabular data, seeing as this is core to your business. 

Karim: There are three broad areas we’re investing in. One is AI and servicing customers within our core product. Broadly amongst our primary customers, which are people in the finance function as well as employees, we think a lot about how to bridge the gap between people in the finance function who have context around policies, and employees who consistently expense but haven't memorized the expense policy, or can’t  easily recollect IRS rules. We want to make this often disjointed back-and-forth seamless, so it doesn’t come across as constant policing. 

Instead of asking users to fill out something starting from scratch, or requesting employees to look through a dropdown to find a category, we offer hints about what it could be, then ask for confirmation so we can map to the right category. 

Is that tied to lots of documents and data flowing through your system or something quite different?

Karim: Absolutely. We get a lot of context about the transaction, and what it could be, from the receipt. We also get some context from the metadata in some cases, IP address, whether you're traveling abroad and a handful of other data we’re leveraging to figure out key questions. We also spend a good amount of time figuring out if a transaction is in or out of policy, based on the information on the receipt.

‍Nik: Traditional finance generally dealt with structured data. LLMs allow us to broaden the dataset massively. Suddenly we have access to data that otherwise is difficult to process. Our engineering team and applied AI team build many solutions in-house, and we vertically integrate to solve these problems. 

One example of this is our universal search copilot. You can ask questions like “What was my highest SaaS spend over the past 60 days?” Previously, we had to select some filters on a table. Now there’s actually an interactive experience within the UI. 

Another use case is oriented around dealing with messy data. In our procurement product, we give customers an option to upload a contract. Most SaaS companies don't have two contracts that look the same – it’s exceptionally messy. Seat intelligence quickly becomes complex if you're trying to ask, "Well, how many seats have we bought and what were the terms of the contract – did we get a special discount?" AI allows us to digest, make intelligent decisions and provide insights we’re already working on, such that we can gain confidence over time, and fully automate this process. 

The other category of embedded AI are things that just happen below the surface that you don't notice. When you swipe your Ramp card, the amount is deterministic, so you actually don't need AI.

Our authorized process that approves and declines transactions leverages traditional algorithms – AI isn’t present. After the transaction is made, we ask you to upload a receipt and type a memo to map it to an accounting category. The biggest friction in spend management products is the user experience. We have a somewhat satirical call in a program inside Ramp called “zero clicks given.” It describes the experience of an employee with spend management. Employees don't give a click about the finance function of a company. Given this, why ask employees to become an accounting expert and select the proper categorization? 

These are the two instances where interactions are happening behind the scenes and we can automate a workflow that otherwise is a big burden for a persona that doesn’t care about or have strong expertise around finance. We also want to create an exceptional experience for finance teams –  they’re using Ramp in their day to day and are the key decision makers with respect to financial software procurement. 

Fast forward to models, infrastructure and data. What is your philosophy around models? 

Karim: It’s hybrid – there are use cases where it makes a lot more sense for us to train our own special purpose models because we need them to be significantly cheaper and have better control over latency. But there are other cases where it's just still very general so a special purpose model doesn't make sense. 

One thing to note is that there are a few key functions you need to have in order to leverage someone else’s model and integrate it into your stack. We do a lot of this work internally. 

Anything else you’d like to comment on as it pertains to Ramp’s infrastructure decisions?

Karim: There are many efforts going into foundation models as well as the infrastructure layer. There aren’t enough efforts going into the application layer. We’re focusing on application opportunities, as there's a significant competitive advantage we can build here. Figuring out the right interface and innovating around how you interact with AI is still very under-invested in specifically.

We are avid buyers and testers of a lot of different tools. We like to move fast and edit quickly. The advantage when you move faster is you can try a ton and be flexible with respect to the infrastructure and foundation model layer.  

Nik: It’s fascinating to have two teams in the company – one that does traditional finance model training and another is working on language models and AI. I'm eager to see if and when these tracks converge, as it seems fairly inevitable to me. The first group is shifting from boosted trees to neural nets in the fraud space. The data and metadata we’re collecting is quite robust – enriched transactions will allow fintech companies like Ramp to make more precise decisions with deep learning, rather than traditional number-based models.

Let’s shift to talent. I’ve been astounded by the caliber of your team – Ramp is quite possibly one of the strongest talent magnets in NY.  How did you establish such a strong engineering culture? What has it been like building the core team in New York?

Karim: My relationship with Nik very much revolves around this – it’s probably the point that we most strongly bonded over. The ambition to build the best engineering team in New York was an early, articulated goal. There were few examples of true tech-forward New York companies – perhaps MongoDB, CockroachDB and Datadog stood out. Given the fintech focus, we optimized for NY. Candidly it was funny to us that Stripe landed in California. 

When Eric and I were first raising for Paribus, many investors recommended we move to SF as there would be unfair talent arbitrage. One day, we had a brilliant engineer reach out – his resume was sparse, but highlighted that he had built the second highest grossing app on the App Store. I asked him why he was interested in Paribus and he quickly conveyed that he wanted to be in NY and we were the only company solving technically challenging problems. 

His comment left a lasting impression – since then, we’ve always focused on cementing our brand as the best tech-forward company to work at in NY. We invested heavily in establishing strong talent pipelines coming out of universities as well. We’ve been committed from the outset to training our engineers – they don’t necessarily need a ton of experience to be impactful, we just need to attract contributors who are exceptionally hungry. 

Walk us through what it has been like to scale the engineering team culture – how do you keep the bar high? 

Nik: Since the outset, senior leadership has been aligned around the fact that we cannot mess up our culture. I always contemplate how we can preserve and scale ownership mentality, especially since our engineering organization is nearing 150 individuals. 

Are you creating no management tier and optimizing for a very flat structure? Are you also creating horizontal teams for cloud ops, databases, security, etc. 

Nik: Yeah, there are some horizontal teams. We strongly believe that some of the best platforms are built when you solve customer problems. We didn’t create a payment platform team until the third payment product. We just had reimbursement and bill pay products fully on the same stack. When a product becomes big enough and demands more attention, we’ll focus more engineering talent towards it.  

Perhaps you can also speak to the culture around people management. 

Nik: We’re currently going through a calibration cycle and figuring out how to manage the team at scale. Karim and I actually sat down a couple of months ago and wrote down what is important to us in terms of culture. There are obviously company cultural values, but we wrote down what is key to the archetypal Ramp engineer. It's primarily customer obsession, strong ownership mentality, and eagerness to build and ship fast. We don't hire people to push paper and do a lot of management. Everybody has to build. I consider my role as an inverted pyramid – head of engineering is a support role with an emphasis on developing talent. The people who make the calls are actually leaf nodes, or the most junior engineers who run the product. 

When you shift from “software engineer” to “senior engineer” in title, the key criteria is: are you enabling others to be more effective? What have you built that made the rest of your team execute faster? And not just frameworks for the sake of frameworks, but cutting down the crud or playing with code-gen or some AI tools to make your data migrations faster. 

I’d love to conclude with some forward-looking projections over the next three to five years – how will Ramp evolve and unlock new opportunities? 

Karim: It’s rare as a company to have one incredible product that has strong product/market fit. We've found clear PMF and are now hyper-obsessed with not only scaling our core product, but pushing to add complimentary products as well. Prioritization is challenging. 

We try to take into account what customers are most excited about, but always weigh this against where we have clarity of engineering/build execution. I religiously sit down and stack rank priorities. 

‍

‍

Thank you to Karim and Nik for their insight and candor (and to the entire Ramp team for being outstanding co-hosts and partners). While we’ve enjoyed fireside chats with AI leaders at companies ranging from seed-stage to NVIDIA, our evening with Ramp provided a window into a company at a uniquely exciting moment. As we’ve read, the Ramp team has always been guided by a clear, simple (but far from easy) mission, and their discipline in approaching each leg of the climb has been remarkable. Among enterprise platforms, Ramp is in rare company in terms of empowering specialists and everyday employees alike, and delivering both a delightful experience and extensive, quantifiable value. We are privileged to be investors, and to work with one of the most focused, innovative, and collaborative teams we’ve encountered.