AI Minds #057 | Ram Venkataraman, CTO & Co-Founder at Sei AI

About this episode

Ram Venkataraman, CTO & Co-Founder at Sei AI, is revolutionizing AI-driven compliance for financial institutions. Sei AI builds cutting-edge, compliant AI agents that transform customer interactions while optimizing costs.

Their AI-powered voice agents automate support, sales, and activation/reminder calls—boosting customer satisfaction and revenue without increasing contact center expenses. Meanwhile, their QA agent ensures compliance by monitoring customer conversations for regulatory and policy violations, while also extracting deep customer insights.

Sei AI is already trusted by publicly listed and enterprise financial institutions across the US, UK, and APAC, setting a new standard for AI-driven compliance and customer engagement in the financial sector.

Listen to the episode on Spotify, Apple Podcast, Podcast addicts, Castbox. You can also watch this episode on YouTube.

In this episode of the AI Minds Podcast, Ram Venkataraman, CTO & Co-Founder of Sei AI, shares his journey of building AI-driven solutions that transform compliance and customer engagement in financial institutions.

Ram discusses his transition from leading roles at Amazon, PayPal, and Transferwise to launching Sei AI, driven by a passion for autonomy and innovation. He recounts meeting his co-founder, Pranay, through the On Deck program, which led them to Y Combinator with just an idea in hand.

Originally focused on B2B payment ledgers, they pivoted to tackling compliance challenges using AI—helping financial institutions enhance marketing compliance and customer support quality assurance. Ram explains how Sei AI’s voice and QA agents outperform traditional monitoring methods, delivering higher accuracy and efficiency in regulated industries.

The conversation dives into the challenges of voice AI, the hurdles of scaling from POC to production, and the intricacies of building compliant AI-powered voice bots. Ram also shares insights on AI pricing strategies, deployment models, and the balance between deterministic and probabilistic workflows for enterprise AI adoption.

Throughout the episode, Ram highlights Sei AI’s approach to integrating AI seamlessly into compliance and customer support, leveraging Deepgram and other cutting-edge technologies to drive efficiency and regulatory adherence.

Show Notes:

00:00 "AI Minds Podcast: AI First Companies"

05:39 Startup Challenges and Experimentation

08:03 Ensuring Ethical Advertising Compliance

10:55 Voice AI Boom and Launch

13:07 Voice AI: Scaling and Challenges

16:22 Enterprise Model Deployment Options

19:43 AI Readiness and Implementation Strategy

More Quotes from Ram:

Transcript:

Demetrios:

Welcome back everyone to another AI Minds podcast. This is a podcast where we explore the companies of tomorrow being built AI First. I am your host, Demetrios, and this episode, like every other episode, is brought to you by Deepgram. The number one speech to text and text to speech API on the Internet today, trusted by the world's top enterprises, conversational AI leaders and startups. Some of these folks you may have heard of like Spotify, Twilio, NASA and Citibank. I have the pleasure of sitting down with Ram, the CTO and co founder of Sei Ai. How you doing, man?

Ram Venkataraman:

Doing great, dude.

Demetrios:

All right, well, I know you've had quite an illustrious history in the financial space. What were you doing in these first companies you worked for Transferwise, or WISE as they call it these days. I'm a user of Wise and I started it back when it was Transferwise. You also worked for a few other financial companies. What were you doing there?

Ram Venkataraman:

It's been a long journey. I graduated sort of in 2009. That's like too far back in the history now. Worked for a couple of companies like Amazon and stuff. Then they wanted to get into financial services. They went to PayPal as part of the team. This is right about in 2013.

Ram Venkataraman:

2012. 2013, when PayPal was migrating away from like a big monolith to like microservices. That was all the. The charm those days. Like APIs, as part of a team that built the APIs for peer to peer money transfer and helped it launch across the world. So with PayPal for like three, three and a half years. And I went to Singapore after that.

Ram Venkataraman:

I was in Transferwise where I was part of a team which helped Transvise launch in Australia and New Zealand. Did a bunch of engineering and product work there and came back to India, started this company.

Demetrios:

And when you were at Transferwise, I imagine you didn't get your teeth into something as big of a project as migrating from a monolith to API services.

Ram Venkataraman:

Thankfully not though. There were similar issues in Transferwise as to that point. But as part of this, a regional team where the goal is not to undertake massive technology changes, but more of like launching Transo in new regions and get the revenue out basically faster.

Demetrios:

So you sound like you were on the engineering side of the house.

Ram Venkataraman:

Yes, mostly engineering, yes.

Demetrios:

What made you then decide to try your hand at starting a company?

Ram Venkataraman:

The real reason is probably because I didn't like working for others. I wanted to do something on my Own. And I felt that life was kicking by without too much fun happening at work. So I wanted to try something on my own. I did try out a bunch of ideas before I met my co founder, Pranay.

Demetrios:

How'd you guys meet?

Ram Venkataraman:

It was complete serendipity. So there is this program called On Deck which more like folks who are interested in starting companies. Like a cohort of sorts. Cohorts plus like blind dates if you will. It was all remote. This was during the second wave of COVID Basically.

Ram Venkataraman:

This was in 2021. We were the plateau on Deck. We saw each other's intros. I wanted to reach out, but I forgot late in the night I went asleep. But he reached out to me that night. And the next day we started chatting. And two days later we were still chatting. Five days later we were still chatting and trying to figure out what to do.

Ram Venkataraman:

And then as we started discussing more areas, we felt that it's like a natural fit. I'm more like an engineering product guy versus my co founder is a lot more on the sales side of things. He has a MBA from Columbia and stuff. So he has actually worked with Travis Kalanick on his post Uber startup.

Demetrios:

Oh, nice.

Ram Venkataraman:

So the cloud kitchens. So he'll expand that in like. Middle east and Southeast Asia basically. So our skills were like perfectly they complimenting each other. So we started jamming a few ideas. We both wanted to get into ic. The goal was that if we don't get into ic, let's rethink if we have to do something on our own.

Ram Venkataraman:

And we applied to ic and in a week we got confirmation that we are both in. This was in July, August 2021. So the batch is supposed to start in the month of Jan, 2022. But we got in very early with just an idea. No product, no customers, nothing. So we were building something like a ledger for B2B payments at that point.

Demetrios:

Which makes sense. That was the space you knew.

Ram Venkataraman:

And in the next six months we got a few customers interested, signed Lois and stuff. But the more we talked to companies, the more we felt that this was like a push product rather than a pull. And we didn't get enough conviction on that. And after that YC happened. We experimented with a bunch of ideas. Basically none of them went anywhere. The worst thing that you can probably expect not to happen in a startup is you can die, but you have to know about it soon.

Ram Venkataraman:

The slow death is what is really frustrating. We are having A bunch of ideas which were just slow death ideas basically. And we are fooling around for a year, year and a half and then GPT happened. And I have seen some of the problems in regulated financial institutions. In PayPal and Transfer Wise like complaints is the core of everything that a regulated financial institution has to take care of. Whether you're like launching new products or like launching in new countries, whatever. Yes.

Ram Venkataraman:

We try to take a look at complaints in a broad perspective. We talked to I think around 50 to 60 chief complaints officers across the world. Try to synthesize what are the problems that they have. We nailed about two core problems that they had which is also a easier wedge into building the broader compliance space. Basically the wedge that we ended up was using AI to solve for marketing compliance. For example, this was in 2023 when CFPB was still there. I mean it's still there, but Elon Musk might have different ideas.

Ram Venkataraman:

So CFPB was very active. Lots of enforcement actions, lots of activity around. CFPB is the Consumer Financial Protection Bureau which is regulating all the financial institutions. One of the core tenets is that when you're marketing or advertising a financial product, it has to UDAP as the regulation. Basically like unfair deceptive action practices.

Demetrios:

Yeah.

Ram Venkataraman:

You should not advertise in a way that it's unfair, abusing or like deceptive to the end consumer. There's tons of enforcement actions that were coming in and this is a product built for that particular use case which is that when you're advertising something on the Internet, whether you are a bank or a banking as a service player or a fintech, you have to be compliant, so we started off with that. We had a few customers there. We always wanted to use this as a wedge into the broader compliance for customer support product, which is when a customer support conversation happens. You as a regulated financial institution, typically you do quality assurance of all of those calls, emails, chats, to ensure that the agents, the customer support agents are not saying something that is regulatory, there is a problem, or even your internal policies, anything outside of that, it's a big problem. Typically in big enterprises they do sampling, they'll pick 5 to 10% of the call. 10% is probably on the higher end.

Ram Venkataraman:

What you have heard typically is like 3 to 5% on average of all.

Demetrios:

The calls that are. How many calls are these big financial institutions getting per day?

Ram Venkataraman:

We're talking thousands of customers that we are working with, they spend $100 million in customer support alone. That's the overall cost on customer support. So obviously there is going to be some X multiple of the number of calls or chats that happen. It's not humanly possible. And this company has like 60 to 70 quality assurance folks whose only responsibility is to listen to a sample of these calls, chats or emails, figure out where the agent is like screwing up. Ideal fit for AI. No need for sampling.

Ram Venkataraman:

100% use a tool like Deepgram to get speech to text. Get the text, run it through an LLM, ingest a customer's policy, use rag, use prompts, fine tune models, whatever. Give them like a report saying that hey out of these 100% of all of your calls, 2% have this issue, 3% have some other issue, 5% this where the agent violated this regulation, So we always wanted to use marketing complaints to be a wedge to do the customer support quality assurance basically. And that's what we did. Like mid to late last year we launched the QA for customer support product. We already worked with a bunch of fintechs, a couple of listed companies and like CDC plus financial institutions are using us.

Ram Venkataraman:

But then the interesting thing that we saw is voice AI just dramatically boomed in the last six months. I knew that this is going to be hot, but I didn't expect this to be so hot so fast. But the good thing for us is that it sort of positioned us in a nice way to launch voice AI agents which we've already launched and working with a few customers that since we started with the compliance first background, since we started with quality assurance for your human customer support agents, we can ingest your policies. We know what those regulations are and things like that. It is sort of easy. With platforms like Deepgram, ElevenLabs or whatever, it's very easy to build a voice bot. But with a background in complaints.

Ram Venkataraman:

We believe that there is a particular niche of building compliant voice bots and selling to regulated financial institutions only who'd care a lot about not just about cost reduction with bots, but more about is the bot saying things that is like simply not true or it's about saying things that are like regulatory violations basically. For example, bought using like abusive language or like deceptive practices, whatever. Right?

Demetrios:

Yeah.

Ram Venkataraman:

Or it's not displaying the, or it's not quoting the disclosures properly. For example, if you're a real estate like a mortgage company, there's a Bunch of disclosures that you have to sort of say so since we have come in from that background versus others who just wanted to build voice AI agents first and then I can add complaints which is a valid approach. But for us since we started with compliance we have put the right guardrails in place already. It was just very easy using tools like Deepgram and LLMs to build a voice AI agent on top of all of our complaints knowledge and just launch it.

Demetrios:

So what do you think are some unique challenges that you've run into on.

Ram Venkataraman:

The voice AI side? on the voice AI side like I said, it's very easy to. It's true for AI in general. It's very easy to build a poc, get a flashy demo out, but the devil is in the details, That's always a challenge like how do we get a PoC ready AI app work great in production. But I'm not particularly worried about that because one on the knowledge side we have enough knowledge about compliance and working with companies across like mortgage buy, no pay, later lending and stuff so we can kind of understand the patterns of what companies look for in AI agent. And on the platform side with I said, companies like Deepgram, LLMs and Stuff, it is very easy and I can trust Deepgram to sort of handle scale automatically. So some of the typical questions that I used to have as a tech founder can this scale if I make like 100,000 calls in a day, will it work? Like I'm like 95, 96% sure that it'll work because Deepgram is built off of so many startups, enterprise customers working off, the cloud, the auto scaling, the pattern should have figured out and like WebRTC and the real time latency reduction and things like that. So I'm like less worried about the tech infra side but I'm much more worried about sort of getting the right customers growing with them moving very fast.

Ram Venkataraman:

That's always a challenge. So that's the bigger worry. Not anything on the tech.

Demetrios:

How do you look at pricing? Because I know that is something that is on founders minds and also on companies minds. You want the companies to reach for you like you're a toothbrush, That you want to have it be something healthy. But if someone knows that every time they're using you they're going to get charged per instance or per outcome or per request, then maybe that's going to make them not rely on you as heavily but on your side every time you're Making a call to LLM or to Deepgram or to LLM, whatever it is, you're paying for that. So if you're not passing it on to the customer, then you're essentially eating the profit. And I find that can be very confusing for founders to wrestle with.

Ram Venkataraman:

So this is the new age of pricing. So with SaaS, typically you charge the fixed fee or charge by the seats or whatever. And the marginal cost of serving another user or another set of scale workloads is very minimal versus with AI, it's all variable cost. So which makes it tricky to sort of calculate the prices upfront.

Ram Venkataraman:

But on the flip side, if you are working with enterprises and if you can expect good enough scale, you can sort of take multiple approaches. One of the approaches, the simplest approach is what Deepgram already supports, that we could get into an enterprise plan and I can deploy it in house. There will still be some variable costs involved, but it's not as variable as like number of calls or the number of minutes of calls that I have every month. It sort of amortizes the costs. that is definitely an option deploying the models in house will reduce the amount, will amortize the variable cost into fixed cost. But more than that, I think we are also getting into an age of outcome based pricing, which is the number of tickets that I successfully deflect.

Demetrios:

Because that's really what I have heard people compare it against. If you are comparing it against SaaS, then you're inevitably going to be a more expensive option. But if you're comparing it against the time of a human to do that same task, then it's going to be a much cheaper option.

Ram Venkataraman:

The only thing that will matter there is how good your product is in terms of accuracy.

Demetrios:

Yeah.

Ram Venkataraman:

That you don't want to do. For example, you don't want to do outcome based pricing if you're not sure that the product performs better than 95% accuracy.

Demetrios:

Right.

Ram Venkataraman:

So I believe that as models improve with the new age reasoning models, the accuracy is sort of like increasing. The costs are also coming down. I believe that more and more companies would want to experiment with outcome based pricing, but I also don't know if it is there yet that you can base off of your entire business model. Maybe Sierra can do. They've raised ton of money. They can afford to experiment. But not a seed stage startup at this point.

Demetrios:

When you are getting the agents inside of the companies and they're fielding support questions how much of the workflows that these agents are doing are deterministic, with some LLM calls involved versus the agent has free range. And the whole thing is the agent deciding on what it does.

Ram Venkataraman:

So all of the use cases that we have deployed, it's still not there yet, but I think it'll get there. But all of the use cases that we have deployed is very constrained. You also want that kind of predictability, because LLMs by default are like stochastic. You cannot predict customers, especially the enterprise customers that we are working with. You need to give them predictability. The moment you talk about probability, probabilistic nature of LLMs, they are already turned off. So the way to get predictability is to pick use cases.

Ram Venkataraman:

I know that when I work with a customer, I will look at some of the conversations that they have had in the last few months. I know the patterns of conversations that they typically have. And internally, we have something called AI readiness scorecard, which is that I look at use cases, I look at how the agents respond, and figure out what percentage of the use cases are ready for AI currently with minimal lift, both from our side and our customer side. If you can provide value based on the set of use cases that we hand pick, then we provide that value upfront. Obviously, the pricing will be very small. They're happy. We roll out to 100% of that particular use case, and when it sits outside of the use case, it will deflect to a human. Then I'll pick the next use case in my scorecard.

Demetrios:

That makes sense. So then you go after the lowest hanging fruit and you get it done. Well, this has been awesome, Ram. I appreciate you coming on here and talking with me about it.