Vellum AI: Akash's $5M Journey

Jake Aaron Villarreal: Welcome to From the Ground Up, the podcast where we delve deep into the inspiring stories of entrepreneurs and their journey to build successful startups. I'm your host, Jake Aaron Villarreal, and in each episode, I sit down with the founders to learn about their experiences, the challenges they face, the lessons they've learned, and the insights they've gained while turning their dreams into reality.

I read about Vellum in a TechCrunch article titled, "Prompt Engineering Startup Raises $5 Million as Demand for Generative AI Services Scales." But what we learned is that prompt engineering is just a small part of their platform, which helps engineers build LLM applications faster. Listen to this fascinating story of Akash Sharma, who graduated from UC Berkeley, spent five years at McKinsey, quit his job at a startup to launch Vellum AI as a founder and CEO. He enrolled in Y Combinator and within five months got funding in today's market. That's amazing. Akash, to his credit, worked on the ChatGPT application since beta was released in 2020 and was clear to him and his co-founders that there was a need to build a platform to bring software engineering best practices to the development process around LLMs today. They're on a mission to provide the tooling and knowledge to help people with taking LLMs to production.

Akash, I guess before we dive in, talk to us a little bit about kind of you and where you're from originally.

Akash Sharma: Yeah. Um, so I grew up in India and I did high school there in Mumbai, India. I wrapped up at an international school out there, Dhirubhai Ambani School, that was great. And uh, when I wrapped up from high school, I decided to move to the US for my undergrad. Uh, applied to a bunch of schools and was really excited about the prospect of going to UC Berkeley. Uh, I was navigating the major uh maze in the beginning, I wasn't sure like what I wanted to do, so I started with mechanical engineering, doubled with computer science a little bit, doubled in material science and engineering. But ultimately settled in on industrial engineering and business administration as my two majors. So the industrial engineering degree had a component of operations research too, so things like supply chain planning, um like scheduling for call centers, more like optimization problems. It was really fun. And I doubled that up with a traditional business administration degree of things like accounting, finance, strategy. Uh, so coming through, like doing that was a great experience and like, just both the mathematical side of business and the, like, the business side of business, which is just all the things that you need to keep in mind as you're running a company or advising that large corporations. So that was kind of the journey from uh high school in Mumbai all the way to college at UC Berkeley.

Jake Aaron Villarreal: That's great. How did you get into the startup world?

Akash Sharma: Uh, sat at Dover. I had a very interesting like uh segue into it. So when I graduated from Berkeley, I wasn't exactly sure what I wanted to do, as with most college students. I wasn't that set on the startup world. So I was like, "Okay, uh let's join a management consulting firm." And uh McKinsey was the best management consulting firm, like it still is. I, oh, I was like, "I'll go there, I'll learn a ton about businesses, I'll learn a ton about technology, I'll learn about software, uh, and just how business is done at scale." So I joined in at the Valley office, worked with tech companies across all sizes ranging from Series B companies all the way to a large massive public companies in all sorts of functions, like sales and marketing, um, strategy, finance. Just got a very like broad functional understanding of how software companies are run.

And there are some um like small experiences I had there where I saw just the power of a good, a well-run B2B SaaS business. Happy to go into some more examples there, but I really, I found that very interesting, like how when you create a good B2B SaaS business, or you can have positive net dollar retention, you can like really improve the productivity of the customer base that you're solving for. Uh, that was a very inspiring moment and I was like, "Okay, I should have, like I should have my own B2B software company at some point where we can help other customers for, with the problems that they're facing." Um, that's where I got the idea. And, but after having spent five years at McKinsey, uh which is a very big company and not not really a startup, I was like "I'm not ready to start my own thing yet," so I decided to join a startup, uh, let's say, startup called Dover, YC Summer '19 company, had an amazing experience there. Uh did pretty much every non-technical function there, like non-engineering function there. Uh learned a ton and then decided to uh launch Vellum at the beginning of this year.

Jake Aaron Villarreal: Got it. That's great. You know, it's uh, it's a moment in time where everyone has to face where you're gonna either start your own company or go work for another company. What gave you the courage to take that step to go out, to quit your job, for your founders to buy in, and also quit their job, and to start Vellum?

Akash Sharma: Yeah, I think everyone has different motivations, everyone is like, uh, one of the things that YC helped us with was thinking through like what our intrinsic motivations are on why we want to start the company. And I sometimes view my life as a uh regret minimization function kind of thing. So whenever I'm uh making a decision about what I want to do with my lives, uh it's, it's happened maybe two or three times like as big pivotal decisions, I always think about like "if I make this decision, five years from now, if I look back at this decision five years from now, and even if, and no matter how it goes, would I regret the decision or not?"

Um, so an example of that would be when I was an Engagement Manager at McKinsey, uh doing fairly well, I would, I was on track to make Partner there, which would have been a pretty successful career outcome coming out of, uh, being a Partner by now. Um, just living very decent lives. But then I asked myself that if I get there, would I be happy with the decision I made, uh and that I didn't even try going for a startup? Um, and my answer was no, I wouldn't be happy. So I decided to just take the leap and see what happens. And it's not like, uh, I don't think that any decision is irreversible, right? I hope that the startup keeps growing and is, is a generational defining company, that's what we're shooting towards. But uh, if it doesn't, then I can always do something else in the future. So it's not like, uh, time is finite, yes, but it's also very long, so we can always like do something else in the future.

Jake Aaron Villarreal: Yeah, that's great. Um, you know, when you take an idea and you bring it to market, you typically go out and you build it, you try and get customers to adopt it, and then you get funding and you go. In Y Combinator, it's a little bit different. Walk me through what they have you do with your idea first before you actually start building.

Akash Sharma: Yeah. So we applied to Y Combinator with just an idea, um, and we did not have much time to validate the idea, and we, we got in. Um, and it was very, very nascent series that, I of this idea. So the first thing they said was, uh, a concept that we still keep in mind back to the date, is called "Sell before building." Uh and it was fantastic. They basically were like, "Don't write a single line of code." Uh because we're all technologists, like we want to build products, uh we really enjoy doing that. Um, but they said, "Don't write a single line of code. Uh convince people to buy your product which doesn't exist yet, from a company that it, that is no brand, no name, nothing. Uh convince them to buy it for at least $500, ideally $1,000 a month. Uh because their philosophy was you need to be able to identify a deep enough pain point or a problem that people are willing to take a bet on you, or us people, and also just take a bet on someone solving this problem for them. Because only then an idea has legs to uh become a big company."

And that's what we did. It was a, yeah, I remember like in the early days of our, yeah, we interviewed about 100 people with no preconceived notions. Like, obviously we had preconceived notions, that's where we came up with the idea, we had some like, uh, we had some concept of what we want to do, but we put that aside. Just like doing very user research, figuring out like what are the top three problems for you as an individual, for your department, for your company? Um and uh then from all those 100 or so user interviews, we feel, we grouped them into like what company sizes they were, how keen were they to solve like to... uh, how likely were they to buy something if we were to create something in the space? And we just did some grouping or synthesized all the findings and came up with a one-pager of like, "this is what we're going to build." Then went back to them and said, "Okay, uh are you ready to buy this without us having written a single line of code?" Then we got there and then that's how the idea kind of started.

Jake Aaron Villarreal: I like it. I think everyone should take note of that where you don't have to write any code, you have an idea, you sell it first, you get the buy-in, and then you build it. It's I think the reverse of what traditional VC which is you build something, you go out to market, you hope customers come, if they don't, you spend a lot of money and you have to pivot and reintroduce it to different markets. So I just think it's a, it's a good approach. I don't think it's maybe as common as people would think, but that's just an incredible experience. What inspired you uh to first even get into ChatGPT? I mean, you're talking 2020 when really no one really even knew what that was. Maybe on the engineering side you did, but how did you get interested in ChatGPT initially?

Akash Sharma: Yeah. So uh basically me and my co-founders, Sid and Noah, like all three of us were, uh, we just saw OpenAI come out with GPT-3 back in the day and uh it was magical. Like uh just if they think about the world before GPT and the world after GPT, the difference that happened was unreal because you write a little bit of text, like around basically and the problems can get very complex, I agree. We write a little bit of text and there's a machine out there that just gives you like something uh from nowhere which is, which is fascinating. So when we saw this happen and admittedly the technology was a lot more crude and the responses weren't very, uh, weren't great in the beginning when GPT-3 came out. Uh we were like, "This thing has so much potential."

So we just started uh building applications which were relevant to our previous company, Dover. Dover is a recruiting software company, also a YC company. And we built, our first application was a job description, JD writer, chapter description writer. So uh you, we would just give like input parameters like uh what the requirements of the role were, what the company description was like, just in a few few lines, and then this would create a job description for you. Great. That, and that created a lot of growth for Dover because it just got like lots of free signups. People are like "Wow, this is cool." Uh, that was one.

Second one was an email classifier. So we were also handling lots of outbound emails for candidates. Um so figuring out whether an income, whether a candidate is interested, not interested, or they are kind of lukewarm, maybe like passive or say warm kind of thing. Um, we had email generator obviously, like creating personalized emails for candidates. So yeah. And I think one thing that's interesting that I mentioned which I just glossed over right now is that the classification use case, the classifier is actually also an example of of LLMs, or large language models, producing results which are important to businesses. Classifiers are actually like very important because then they can, you can use a good classifier to, to determine what to do next in your application. Um, and LLMs can also be used to classify text into some category that you ask it to classify on. So we saw all this happen and we were building these applications, we're like, "Okay, this is interesting, like we should do something here."

Jake Aaron Villarreal: Yeah, that's great. For people that don't know what LLMs are, walk us through what they are today and where are they prevalent in the marketplace within companies?

Akash Sharma: Yeah. Um so I think the closest example that everyone may have experienced or seen is ChatGPT. Um and LLM, large language model, is basically a model which takes in text as a prompt or an input and gives you a... gives you a response which uh, based on the knowledge that it's been trained on. Uh the usual training cutoff is for OpenAI's around March of 2021 or something. Um, and uh that's that's basically a very, very high level of what LLMs are. But these different companies like OpenAI, Anthropic, Cohere, Google, they all have their own different models. So the same text that goes into one model produces different results.

And you may also have heard of Llama 2, like the open source model from Facebook. There's some other like pretty promising open source models like Falcon 40B built by the UAE government, um and MosaicML's MPT series of models. These models are open source. They can be downloaded and they can be installed on your own infrastructure. And they also perform like in a similar, in a similar way to these other closed source models, right, that's as I call them. The only difference is that you have to host these models yourself. The other ones you can just make a call to an external API and get results back. Now that's kind of how they work... I mean, that's kind of what they are. How they work, it's really hard to know, like what, how it actually works. Uh, but that's where an experimentation platform comes in handy where you can compare how different models perform when you give it similar or slightly different prompts, or even the same prompt.

And companies usually look at the quality of these responses, the latency... latency is very important because you want to make sure that your responses come in a timely fashion. So people look at latency, quality, and also cost. The cost profiles of all these models are quite different. The most common application that has existed since GPT-3 came out was more on the email generation case, like marketing copy generation, email generation. There's a lot of companies out there that do that. I think Jasper is the tier leader in that space. They built up, they built this company pretty like, right when GPT-3 came out and they grew quite a bit. So that's the most common use case.

But as, as people are discovering like everything that LLMs can do, they're looking at classification tasks. They're looking at data extraction tasks. Data extraction is so interesting because think about um having like lots of different PDF files which have lots of different texts and stuff out there. Um, existing OCR technology is, is pretty decent at extracting fields. But OCR technology doesn't really work when there's uh, like the text is not standardized or the format changes a little bit or there's spelling mistakes. But LLMs can get through all this type... LLMs can find, can make structured data out of unstructured data, which is uh very cool for lots of business applications because there's oftentimes there's humans doing this data entry process which a lot of which can be done by LLM. So classification tasks are cool, generative like email generation tasks are always there, um and um also data extraction is pretty interesting to me.

Jake Aaron Villarreal: That's great. So it's all about AI and helping automate the process that we typically do today, whether it's generating emails or other tasks within your product. Just kind of walk us through what inspired you to take your product Vellum, your company, and and find the niche in the marketplace and really build around that. Like what, what problem are you helping solve designers, developers, people in general around data or around applications in the AI space?

Akash Sharma: Yeah. Um so it's very easy to... I guess the main realization we had was that it's pretty easy to come up with a prototype of something that demo is valid, something that just like looks cool as an application, like an email generator for example. But well, the bar to put it in production, like for an actual business use case and having the human not do that task anymore, it's a bit higher. Like you want it to perform well in all sorts of bad cases. So that's the realization that we had and we continue to experience that when we talk to our customers on a day-to-day basis, that people want to be able to test these models at scale before putting them in production. So the testing process and iteration process is, is pretty important.

And that's what we spend a lot of our time on today, where we give companies an environment to look at different models side by side. So GPT-4 versus GPT-3.5, this is Claude 2 versus uh Llama 2, number of models keep, keep increasing. But uh every company has a different threshold on quality, cost, latency, and we help them like find the uh the, the most ideal point in that, in that scale and the quality, cost, and latency scale um for their use case. And they can do testing at scale using our platform. That's a pretty important problem.

And once... so this is all, this is all pre-production. Uh companies also end up creating like pretty complex prompt chains where they have a classifier up front and the classifier results in different actions downstream. We help them like visualize those chains inside our platform and test uh, test substrate with multiple different input values. Um that's something that people, it's hard to visualize a multi-step LLM chain because uh, uh even the first step is hard to visualize. You don't know what the response looks like from the large language model. But then that's why multi-step gets even harder, but it's important, needed for some business applications. We help people with that, with experimenting on that.

And ultimately, like when it's in production, people also need to know like how it's doing, like what the quality looks like. Um if OpenAI is down, can we reroute the traffic to maybe a different OpenAI model or Anthropic, or... yeah. So I think this is like the current set of problems we're solving. There's more future problems that are on the horizon, happy to go there if interesting.

Jake Aaron Villarreal: Yeah, what I'm trying to get at here is where is prompt engineering in the process of your application? So is it solving, is it taking the prompt engineers out of the process where your platform doesn't need them, or is there... does it help give companies guidance with how prompt engineers should use your application or your tools?

Akash Sharma: Yeah, for sure. So prompt engineering is uh kind of essential, like it's the way that people interact with these models. So I think prompt engineering uh continues to be like a pretty dominant way of people interacting with LLMs. So our product helps companies come up with good prompts. And the process of coming out with good prompts is kind of iterative because you need to keep... basically, it's a set of instructions you're giving to the model that uh, and and the model is giving responses to each of those instructions. And when you give the model these instructions, uh you are tweaking the text, you're changing the variables, uh and you're also testing those instructions across multiple different input values or test cases. And the responses that come back, you are trying to validate whether the responses are good or not. I think there's different ways of measuring large language model quality um and it's kind of subjective, like honestly. Especially in non-classification use cases. Classification is the correct answer yes, no, or like multi-class classifier.

But uh the process of prompt engineering is an iterative process where an engineer, maybe a software engineer, maybe a prompt engineer who may not be a technical person, is basically iterating on their prompts to solve the use case that they need to solve. And we are exploring ways to help people like more automatically come up with prompts, uh that's like on the roadmap, but right now we just help customers kind of manually. Because honestly, prompting techniques are also getting uh like new prompting techniques are coming up with newer models. So uh it's we are also learning as we like interact with dozens and hundreds of customers right now.

Jake Aaron Villarreal: That's great. Um when you look at your product today uh and you got it into the first hands of your clients, walk us through how that happened and give us a specific use case of a company that's using your your product today.

Akash Sharma: Um yeah, so I think uh let me give an example without sharing the name of the company. Uh but this is a very like interesting and amazing example of how uh someone is using a platform to build a very powerful large language model application. So what this company is trying to do is take real providers, so real doctors like medical professionals, they're taking real providers and making quote-unquote agents out of these providers. Because they want to have a large language model that basically talks like these providers and gives patients like recommendations and um advice, health advice, medical advice, similar to how a doctor would, that that doctor would give these patients. And obviously there's like HIPAA compliance restrictions, and at some point it can be giving too much medical advice, there's some guard rails we need to place on the model for that.

But in any case, like uh what we're doing here, one of the prompts, so one of the use cases that we're building out with this customer is how do we onboard this patient? So it's a challenge, like how does a provider onboard a patient? So in the provider and the patient onboarding process, we're trying to collect like information like name, email, symptoms, diagnosis, date of first diagnosis, uh and just a lot of fields like that. But if you just have an LLM just ask questions, like go through a checklist and ask questions, that's not fun. Like that's not how a provider would be interacting with a patient. Um so what we do here is uh it's a it's a multi-step process where in the middle, at each step of the chat conversation, we ask uh we we ask the LLM whether there's something relevant in the provider's long-term memory.

So by long-term memory I mean uh just a collection of documents that we've uploaded, that the customer's uploaded into our platform, which uh which describe like what the provider knows, like what the provider knows. It's like flyers and documents or whatever. So given the chat history so far, is there anything relevant in the provider's long-term memory that we can tell this patient? So more like an educational process. Like if the patient says that "I have this kind of symptom," the provider could be giving some tailored advice based on the provider's proprietary knowledge. Um so we do that check first, then we provide the answer, and then at the end of this chain we also modify for tone. Because each provider has different tones and we kind of set like tone changes. So the message comes, is generated by the LLM, and instead of sounding robotic like um "Thank you for that question. Here's the next question," uh we modify the tone now through another prompt and make it talk as if the provider was talking. So we give, we give the prompt like specific memory from the provider and we also modify the tone and make the conversation natural and flowing.

Jake Aaron Villarreal: That's great. You know, in today's market it's really tough to get funding. You know, we talk to a lot of startups that are, you know, trying to get funding, they've been trying for 12, 6, 18 months. For you, you got funding pretty quick. Walk us through the process of getting funding in today's market because you have to have a real problem you're trying to solve with a good team and ultimately that's just something... believe in what you're building.

Akash Sharma: Yeah, so I'm I'm very lucky to have, well to work with my co-founders and now all the engineers on our team. Like I think we are lucky enough to have a good team and we also are solving a problem that's a top three problem or like a hair on fire problem for a lot of companies. Because uh as you may have seen in many of these uh investor reports that happen, like 10-Q and 10-K reports that happen with other companies, the buzzword of AI keeps happening and everyone's trying to implement AI. Uh but companies still struggle to get like uh high-quality applications out there. And uh it's, it's as I mentioned, it's easy enough to make a prompt and put it in production, but it doesn't work with a lot of edge cases. And that's where the companies know that uh using a platform like ours and getting best practices from us will help them build these better applications that they can actually put in production.

So it's a combination of having a team, having the domain experience and the exact like, actually having built these products for many years before the new boom happened, and also uh identifying a good problem that's a top of mind problem for our customers. Uh we are fortunate enough that both worked and we were able to get good initial traction. After those hundred interviews that I spoke to you about, uh got good initial traction, uh continued to have like good ongoing growth. And that's why fundraising uh was an easier process for us compared to most other companies.

Jake Aaron Villarreal: Did you use a specific framework for your pitch deck and how many times did you pitch before you got funded?

Akash Sharma: Um I must have done like over 150 pitch meetings. Um and uh there was... that was like probably like 20, I, I don't, yeah, probably 20 of them were successful, 80 were duds. And that's, that's okay, like, oh, not everyone has to buy into your vision, that's totally fine. Um and in terms of framework, um YC helped us quite a bit and also my prior experience at McKinsey helped me quite a bit because uh, at McKinsey I used to like write presentations all the time and present to clients all the time. So it was uh... YC's advice was "come up with the vertebrae," I was like. Um basically, your your traction so far, uh who your customers are, what problems are you solving for them, uh what the market opportunity looks like, what the upcoming roadmap looks like, and why is this a hair on fire problem? Like five to seven questions like this. Uh and just write those answers on a piece of skateboard before making any slides. Uh make sure those answers are good, convert them to slides, then start practicing. So maybe I practiced with my co-founders for about like 30 dry runs, they asked me all sorts of questions, and then eventually I did about 150 meetings, and that's how the funnel proceeded.

Jake Aaron Villarreal: Got it. How did you go about getting those meetings?

Akash Sharma: Um, so again, luckily I think Y Combinator made it easier. Um well, Y Combinator has consistently had this track record of producing uh multiple unicorns every batch. Um so in the investor community knows that uh the chance to invest in a hot YC startup is uh is a good... at least a chance to look at a hot YC startup is a good thing. So they were willing to, they basically reached out. Um and I was able to schedule calls and only talked to them and were prepared to uh to do our fundraising. And uh we we also, I also got more calls scheduled through some folks I know, uh mostly from Dover, uh our my previous company. And they were able to connect us to more investors and we just got these calls scheduled. And now I send them up for some of these investors too on an ongoing basis.

Jake Aaron Villarreal: Yeah, it's important once you get the funding to keep that relationship good with those investors and the consistency of keeping them updated on how things are going and your progress and all that, I'm sure you're learning about that too. How did you um find your partners?

Akash Sharma: Yeah, so Sid, Noah, and I all were already employees at Dover, uh YC Summer '19 company. And uh we worked again on multiple projects. Sid and Noah actually also had experience in the MLOps industry. So MLOps is analogous to LLMOps. Right now we are in LLMOps. MLOps is analogous but just for a different buyer profile and like different kinds of problems, but there's similar things like monitoring and versioning. So they are both MIT engineers, and uh we when we got together, uh we we all like joined early because we all of us had a inkling to start our own companies like independently. But there was a 50 to 90% chance that each of us would start our own companies, and then we decided, we got together, we were like, "Okay, we have very complementary skills. We also have stumbled upon an idea which is likely to take off." Um and we were like, "Okay, we anyways wanted to do this. So we should just take the plunge and take a leap and see what happens and apply to YC on the back of this idea." And YC was again incredibly helpful to the process both to validate the idea um and to help with fundraising at the end.

Jake Aaron Villarreal: What's uh... today, how big is the team? You've got the founders, you got yourself obviously, the three co-founders. What, how big is the team you have and where are they located?

Akash Sharma: Yeah, we've hired four more engineers, um and they're all amazing. I really like working with all of them. Uh my co-founders are moving to San Francisco, they are currently remote, um they're basically wrapping up where they are, they're moving to San Francisco. Um and I'm in New York, and the engineers that we brought on, most of them are in New York. One of them is, one of them is in Toronto, uh he'll be moving to New York in a few months. So we probably have a bi, probably have a bi-coastal presence. Uh it should be fun. I'll be traveling to SF quite a bit. I used to live in SF for many years before, like when I was in McKinsey and even at Dover. Moved to New York a year ago and then I decided to stay, so I might go back at some point.

Jake Aaron Villarreal: Yeah. When you hire people, there's obviously a process to it. And at the same time, you're really building your culture, especially with that foundation team. What are you looking for and what type of culture, what are you looking for the people that you bring on, uh not just technical, but the fit for you? And also what type of culture do you want to build?

Akash Sharma: Yeah, um, I think that's the... yeah, we think about this quite a bit, right? I think it comes down to how much you care about customers. Like, how much do you care about solving a deep problem for your customers, and uh how much do you like want to make sure that they're successful? Because uh it's it's very clear, like if you don't solve a problem for a large enough subset of customers, then the company is just not going to work. So we try to check for that at every step in the interview process. And even now, like we have our uh engineering team handle customer support uh because we won't have customer support folks for a while because the product is still rough at the edges and uh we're still building a lot. So we want to put engineers there in front of customers doing customer support, doing customer success checking calls, maybe even doing sales calls. Because uh then that, that shows that they really care about uh what the user problems are. And then when they find issues with how the product works or it's like unintuitive, they can also go and fix it, which is the beauty of having uh engineering doing customer support in New York. Yes, obviously that's unscalable, but in the early years we want people to set an example of caring for the customer, staying close to them, staying super responsive in, in the product rollout.

Jake Aaron Villarreal: So far you've got your clients, the early adopters really. What, what's an ideal customer for you? So if someone's listening and they're saying a big enterprise company or maybe a small company, what for you is an ideal fit based on where you're at today?

Akash Sharma: Yeah. So uh we've seen more success with engineering departments, um and the buyer is probably a Director of Engineering, VP of Engineering. People who want to bring software engineering best practices into the world of large language models. Things like unit testing, regression testing, end-to-end testing, all that is available on our platform, with a slight twist because large language models are different from uh traditional software engineering. Um so a leader in an engineering organization trying to bring best practices and also improving the productivity of the engineering teams. Because one thing that we hear consistently is that people who use our platform see a three to four time productivity improvement in building these LLM-operated features. Because you can test multiple prompts like pretty quickly, you can have non-technical people enter the development process, uh and the prompts don't necessarily live in your codebase, they live in a software which is meant to host prompts, which is our software. So yeah, that would be the ideal customer profile. In terms of company size, it can range from two-person companies all the way to 10,000-person companies we have, and also all sorts of industries.

Jake Aaron Villarreal: And what's the model, how do you do... is it like a per seat license? Do you, is it, what's the model look like?

Akash Sharma: Yeah, the pricing model is uh kind of based on what features a company ends up using. So we, our platform has three or four different parts, uh depends on what features the company ends up using. And if they use a fine-tuned model or like an open source model that's hosted by us instead of using OpenAI or Anthropic, then there's also a request cost. So there's a base subscription fee based on the number of products and features someone uses, and there's a per request cost if they use a model that's hosted by us. So it, it falls in the mid hundreds to mid thousands per month range uh depending on those factors.

Jake Aaron Villarreal: Great. You know, as you start the company and now you're building here at four, five, six, seven year engineers now. Um, when you continue to scale up, what, what kind of, what are you looking for? What are your projections as a company size? I know you're early stage, you know, five, six months into it, you've got funding. Where, where do you kind of see things going from here?

Akash Sharma: Yeah, so I think it's um, it's it's out here I think we need to establish like deep product-market fit. In the beginning, we're still uh in that phase. Like um, Marc Andreessen has this like famous article on finding product-market fit and he says that it's an experiential thing where like money is piling in your checking account and you can't hire enough sales and customer success people and it's just like the market is pulling you like a whirlwind. Um, we need to get there first. I think we need to understand like, we are definitely at early signs of very promising and we're actively like progressing towards that. So that's step one, like getting to that market pull, like extreme market pull. And uh in parallel, we have some revenue milestones we hit. The revenue milestones, we consistently keep growing to until we get there. And beyond that point, it's like "Okay, how do we make this uh scalable? How do we add a structure on sales, customer success, engineering, and then grow out the company?" So I I would say it's a bit too, a bit premature to think too much about like a financial forecast 18 months out, 24 months out. Uh it's something that we will have to do at some point, uh but right now we're just getting into, getting to a point where we are very confident that we can solve this problem on an ongoing basis for like thousands of companies in the world.

Jake Aaron Villarreal: When you started this company five, six months ago, and it's really for us, you're kind of one of the the first companies that you're so young, but you also got the funding and you got a real problem you're solving. What's been the biggest challenge so far as you start to build this company, as it evolves, that you didn't anticipate when you began?

Akash Sharma: Yeah, a big challenge that I didn't anticipate, it's an ongoing challenge, is that people in... who are trying to like do AI, gonna go do AI, don't really know what they're doing. Like this often, it's not clear. People did a lot of hobbyists, there's a lot of people who are just exploring. So it's hard to think and people also have like very different maturity levels with how how much they know how to use AI and LLMs. So it's really hard to uh get signal from the noise and figure out like uh who is the, like what level of maturity is the person that we are like ideally solving for. We have a hypothesis, but lots of hobbyists come in and trying to filter that out is a little bit difficult because uh I think this technology is still very early in the adoption curve in the enterprise. And there'll be, in the future there will be tens of thousands, hundreds, hundreds of thousands of companies using it, using LLMs at scale, but right now there's a lot fewer. So we're just trying to get that signal from the noise. A lot of noise in the market right now.

Jake Aaron Villarreal: Yeah. You know, you've been on TechCrunch and you've been in different area areas within the media. How do you go about getting publicity about your product versus all the other AI companies out there?

Akash Sharma: Yeah, um good question. So I think, oh, we care a lot about our messaging. I think our messaging is uh, it needs to be like super consistent and solid. Like a pain point that we consistently clear in the market. So develop a platform for building applications on large language models and this platform will provide the tools and best practices needed to uh bring LLMs, like bring high quality production applications. Um so things like improving engineering productivity, um and also just like identifying specific problems. An example of a specific problem is how should different engineers, or engineers and product managers, collaborate on prompts when they're making these prompts together? Uh status quo is that they are collaborating on Notion or word spreadsheets. But we are trying to change that. We're trying to uh have a better prompt collaboration platform as one of our, one of our things in the platform. So I think basically identifying the correct problem and articulating it and then having marketing messaging associated with that is basically what helps us stand out. Um that's, that's something that worked for us and we'll keep doing that as we go, as we keep growing.

Jake Aaron Villarreal: That's great. What's the one-liner value proposition that you have for your company today?

Akash Sharma: Um yeah, it's the developer platform for uh helping companies build production quality LLM applications.

Jake Aaron Villarreal: That's great. Um Akash, it's been incredibly cool to hear your story and understand what you guys are building. We see nothing but growth in the AI space. We have hundreds of companies we're working with today and the question really isn't are they going to continue to grow, it's finding the right people. How do you go about finding people that you feel like are going to be a good fit for your company as you grow?

Akash Sharma: Um yeah, we've uh mostly hired in network so far. So uh we've gotten lucky with the exceptional people who have somewhat expressed interest in working at Vellum and we've either, we knew them personally like extensively for many years um or we knew someone who knew someone and the intermediate person was someone who was mutually trusted by both parties. Um and uh that's why I think hiring people in network and people who really care about solving customer problems um is what we've done so far and we... it's worked well for us and we continue to do that as much as possible. And at some point it'll have to change as we increase the rate at which we hire uh and we'll have a more standardized recruiting process for that.

Jake Aaron Villarreal: Um, is there anything we haven't asked you that you want to share?

Akash Sharma: Um, um no. I think I think this is a pretty good uh overview. Uh maybe one thing that I would add is that uh, um fine-tuning is something I haven't mentioned on this call, but we are very, very bullish on fine-tuning. Uh fine-tuning is a process of taking an open source model like Llama 2 or Anthropic... oh sorry, Llama 2 or uh Falcon 40B or MosaicML and training it for specific tasks. And when you take a smaller model to train for specific tasks, uh the results are almost as good or maybe sometimes even better than closed source models like GPT-3.5 or GPT-4 at way lower cost and way lower latency. So it's just a whole new like different technique, it's not, it's not a prompt-based model anymore. It's a fine-tuned or trained model. So those models are very bullish because at scale, I think in large companies, I think like a Fortune 100 company uh we'll see like thousands of models in production, a mix of OpenAI, Anthropic, and also these fine-tuned open source models. And we hope to be the platform there that facilitates and orchestrates all these models and shows our customers like what the quality is and how they can keep optimizing their quality, cost, and latency.

Jake Aaron Villarreal: That's great. For the engineers out there listening, I'm sure you know exactly what is being shared. And for the businesses out there that are having large amounts of data and large language models in-house, um Vellum sounds like a great product. Not just a product, but a great team. So if they want to know about you Akash or find your company, where should they go?

Akash Sharma: Yeah, uh they should sign up on uh vellum.ai. The company. And uh feel free to email me at akash@vellum.ai. Happy to uh um chat with anyone and say hello.

Jake Aaron Villarreal: Great. Akash, first and foremost, thank you so much for sharing your time with us today and providing your knowledge and interest uh in voicing your opinion about what you're doing to our listeners. And I want to also thank our listeners um deeply for spending your time with us today and hearing Akash's story. And we look forward to Akash, following up with you in the future and getting progress check reports on how things are going, but super excited to hear how it goes. And thank you so much again for your time today.

Akash Sharma: Awesome. Uh, thank you for having me Jake. Okay, have a good one. See ya.

Jake Aaron Villarreal: Before we wrap up, I want to give a big shout out to all the entrepreneurs that are joined to make this podcast possible, and for all the listeners for listening. It means the world to me that you chose to spend your time with us today. I'm your host Jake Aaron Villarreal, signing off for now. We can't wait to connect with you all soon on the next episode. Take care.

This show is sponsored by Match Relevant, a company that helps venture-backed startups find the best people in the market and they do it in three simple steps. First, they sit down with founders to understand their story. Second, they tell their story into multiple candidate channels. And third, they schedule interviews within 48 hours. Find us at matchrelevant.com to learn more about how we do it.  

‍