Anekha Sokhal spent nearly a decade trading energy derivatives at JP Morgan, then came to the US on a Fulbright to study data science at Rice University. Today she's the founder and CEO of Moshi, a Houston-based, VC-backed real-time quant platform for institutional finance. Ahead of her talk at AIE New York, we talked about why she calls Moshi a data company, what it's like to build in Houston, and the growing gap between how easy it is to build with AI and how hard it is to run what you've built.
Moshi, Anekha, and building in Houston
Tell me about Moshi.
Moshi is a real-time quant for institutional finance. We provide real-time analytics for financial traders. Every time you open Moshi, you're going to learn something new because of the volume and quality of data that we watch.
A fun fact about Moshi: it takes a human four seconds to read a sentence. In that same amount of time, Moshi has learned 300 new things. We watch more than 2.5 billion data points, and we grow our internal database every minute.
The value we add is that you're going to learn something new about your market every time you look at Moshi, and you're going to be so much more efficient. The problem traders have is that they're trying to keep up with everything, and everything is changing constantly. When you're a financial investor or trader, what you knew this morning is no longer true by lunchtime. The news is changing, Twitter is changing, Instagram is changing. People are flooded with data. And the problem with that is: What is true? What is relevant to me? What can I trust? That's where we come in. We are that one-source platform.
Typically, you would hire a team of analysts to do that for you. I used to be a trader at JP Morgan, and I could never even hire an intern, so I was doing all of this myself. So I built the platform I wish I had when I was trading. Because I'm also a data scientist, a lot of machine learning and statistics has gone into it. We've built all of our own models.
And tell me about yourself.
I'm Anekha, the founder and CEO of Moshi. I've grown Moshi myself. We're a VC-backed startup based in Houston, and the reason we're based in Houston is that I used to be a trader at JP Morgan. I was there for nearly a decade, trading energy derivatives.
I really just love data. I'm obsessed with data. I got a Fulbright scholarship (England has one Fulbright scholarship for data analytics) and came to Rice University to do my master's in data science. I got really into machine learning there and was published at conferences like ICLR.
When I was interviewing for jobs, a bunch of hedge funds reached out, and they all had this same problem of data synthesis. Instead of building it for one hedge fund and doing very similar work to what I did before my master's, I thought: I'm in the US. There's a huge venture industry. It's such an entrepreneurial country. That's why I came to the US. I really wanted to be part of that ecosystem.
I did an accelerator program while I was at university, then happened to meet a venture capital fund that loved the idea of building Moshi for everyone. And that's where we are today. My story is very intertwined with the company's story.
That sounds like a wild ride. What is it like to build in Houston?
The straightforward answer is that you get to be a big fish in a small pond. But I do feel like you have to keep your connections with other cities alive.
We're backed by one of the biggest venture capital funds in Texas. If someone is coming to Houston and wants to know which AI companies or finance companies are there, we're top of mind for everyone, which is great.
However, we can't ignore that, at least in our situation, AI and finance have other cities that are very important. San Francisco is one. New York is one. London is huge in finance, and so is Singapore. Houston is a very good city, especially for energy. I used to be an energy trader, and a lot of energy clients are here, so it's great in that way. But you also need to be connected to the other cities, because the hype is happening there too. San Francisco for the venture environment and the people. Austin has a great ecosystem. London has a fantastic finance community. And New York, I feel like, is coming up a little more in startups now, and it's obviously huge in finance.
So what is it like building here? There's a fantastic community. Some of the nicest people I've ever met in my life are here, so many helpful people. You can really make a difference here. But there aren't always going to be as many events, and it's not as easy to meet people because it's a very spread-out city. You don't have those shoulder-to-shoulder conversations you'd have in a walking city.
A data company in an AI world
So, what is your personal relationship to AI?
For my undergrad, which was a very long time ago now, I studied MORS (mathematics, operational research, statistics and economics) at the University of Warwick, and it was not a sexy degree. If I told anyone I was studying MORS at Warwick, they were just like, "Oh my God."
Because I loved statistics and data, I went on to trading. It was a very natural transition. Trading is just data and machine learning. And when I was doing it, no one cared. No one thought what I was doing was interesting or sexy.
I remember once meeting a family friend with my best friend, who was studying law. I'd gotten my internship at JP Morgan at the time. The family friend asked me, "What about you? Do you want to be a tutor or a teacher?" Then I looked at my friend and said, "Wow, you're incredible. You're going to be a lawyer."
It's so different now. I graduated in 2014, and now if you say "machine learning" to people, you're blowing their minds.
So my connection is that I was in it before it was cool to be in it. Now we have this separation between machine learning and generative AI. I call Moshi a data company because I love data. We have data scientists. I'm obsessed with data. One thing we do that's very different from most companies is that we do so much data engineering and machine learning, and then generative AI sits on top of all of that. So my relationship is: I want to understand the data, and generative AI is how we communicate it to people in a more human way.
The bar for calling a company an AI company has gotten way higher. It's not a differentiator anymore.
I finished work super late last night, and to decompress I went on Instagram, because it's the most mindless scrolling. Embarrassingly, I was on it for 40 minutes, until I got the timer notification. This was at 1 a.m. In that time, I got 25 ads for AI.
I screenshotted every single one and sent them to Rick, my head of growth, and said, "Rick, we cannot advertise on Instagram." Twenty-five companies in 40 minutes calling themselves AI, AI for finance. It's so easy to generate. How can I trust them? I think that says everything about what it's like to build today.
I'll never forget driving in from the San Francisco airport earlier this year. There was nothing but AI billboards.
There was a time when our investor told us we needed a .ai domain. I don't really believe in all of that, but we got the .ai domain, and now I don't even call us an AI company. I say we're providing a solution, and part of it does have generative AI, as you'd expect. That's an expectation now. If you have an LLM, it's not cool. It's the expectation.
It's the same in media relations. Five years ago AI was unique. Now the best way to get a company coverage is to not talk about its AI.
It's definitely different. Before Moshi, while I was at university, I tried to start another company, which I closed down, called Jewel Vision. My parents are in the jewelry business, and I wanted to give people the ability to try on jewelry in 3D without even needing models.
It was too early. The technology didn't exist. It was 3D computer vision. To this day, I have recruiters asking, "Can you come to Apple's recruiting events? Can you come to Tesla's?" because it's not a skill people have. I tried to build a company doing it for a year.
That was super cool AI. That is pure 3D computer vision. But if you have an LLM agent, to me, that is not AI. There's a difference.
Prompt fatigue and where humans come in
Where are you seeing prompt fatigue show up?
I want to explore how our work is changing and where humans actually play a part. One thing I've found is that people are now just copy-pasting Claude output to each other. So where does that stop?
If I think about the whole workflow of where we are today: people want solutions that make them more efficient, make them better decision makers and integrate their workflows, and they're using generative tools to do that. The questions that come up a lot for us are: How can I trust the output? How can I connect everything? And what about the data? Where are you getting the data from?
The biggest thing we've realized is that the foundational companies are making it feel like it's so easy to build. And it is easy to build. You can get a $200 subscription to any of them and build a dashboard for yourself, and it will be fantastic. I do it for my wedding planning. I have my little dashboard. Now try to make that production-ready. Try to make it run for more than one person. If I want to send that dashboard to my mom, who doesn't have a Claude account, that's the problem I want to talk about.
What happens when you try to take that into an enterprise?
When you try to implement that in an enterprise, you actually need to run the system. And when you need to run the system, you need DevOps, you need error handling, you need to pay other people to run it. That's the bit no LLM is going to tell you upfront.
So yes, you can spend weeks building a dashboard on Claude. Now try to make that dashboard available to the public. Now you need DDoS protection when you host it on a website. Now you have to make sure your API keys are hidden and not in the code. You have to make sure you're not showing your code on the website. You have to get a database. You can't just host it locally. Those are the things that end up wasting people's time.
Instead of managing it yourself, giving that job to a human, or even doing it the old-fashioned way, you've built it yourself on Claude. But now you have to run it. So you're dealing with prompt management, because every time you want to make a change, you wanted to change one thing and you've gone and changed 50 other things. That's what I wanted to talk about: the prompt fatigue, and the realization that I can build really easily, but I cannot run a system as easily.
So where do people come into all of that?
I think companies are still figuring that out. CEOs and CTOs are constantly posting about how, oh, let's fire everyone, because now we only need two or three people to do that job. And I'm sure you've seen the AWS outages and all these other outages. This never used to happen.
What we're seeing now is a confidence everyone has been given by these AI tools: I can build anything I want. But the thing people are forgetting is that every single person has access to the same tool. Now you're going to tell me, "But I've got a different skills.md file, or a different Claude connector, or a different MCP." Yes, I agree, those are different. But you are still pulling and pushing the same data. It's still the same model. If my engineer asks Claude a question, they're going to get the exact same answer, even if the prompt changes a little.
So where do humans come into this? It's the judgment piece. Claude can't build Moshi. Why? Because I had to take 10 years of derivatives trading, and I'm a data scientist as well, so I knew how to build the system. I had people sitting with me when I was 19 or 20 years old at JP Morgan, spending three hours explaining systems to me. I took all of that, and that's what I'm bringing to my agents. A lot of people are not doing that.
I think one thing we're going to come across is that things are going to break more easily, because of the trust people have in these systems. One tool says, give it all your bank details, your credit card details and your email details, and it will go and run everything for you. We don't do any of that internally, and we have engineers from Microsoft and Meta. No one on our team does that. You can ask them to, and they won't. It's similar to the CEOs of all these tech companies who won't let their kids use social media.
I think there's going to be a sort of awakening. We have big enterprise clients who come to us and say, "Why do I need you? I've got an Anthropic subscription." And I say, try to run that same thing for more than one person. You're going to have security issues. You're not going to want to manage it. You're supposed to be trading and watching the markets. You're not supposed to be telling Claude that FOMC is today, not tomorrow.
There's going to be friction for a while. It's like a purge. People will do all the AI, and then they're going to realize: actually, I want to see a spreadsheet that I have to use my brain to check. I want to remember someone's phone number. I want to see a PDF that isn't perfect. I think you just have to let it happen.
The interesting thing for us is that enterprise software is changing. We have clients sending us their prompts. We had a client today who said, "Hey, can you just run this for us?" And I said, that's not what we do. And they said, "Yeah, but I don't want to run it anymore. You guys do a better job. You do all the trust, you do all the things." People are interacting with us like that now.
Did you do it?
We had to build that, because we didn't know we needed to before.
For us, having been a trader for so long, the human brain is so much more advanced than machine learning. The number of connections we can make in real time, the power of the human brain: there is no data analytics that powerful. We're trying to get closer and closer to that human way of thinking, but machines are not there yet. So why would you want a machine to make such important decisions?
Claude is an incredible tool, but it's incredible at summarizing what you tell it to do and explaining the prompt. What it's not great at is making those second- and third-order connections, which is what humans are great at. It's seeing around the corner, being able to say, "This is what I think I should do," or "This is what I don't see." Those gut feelings, those human interactions.
Can you give an example?
A simple one is hiring. So many times LinkedIn will ask, "Do you want us to AI-interview this person?" And I'm like, no. I want to interview this person.
When I was hiring a couple of weeks ago, I ran a simple experiment. I ran every résumé through Claude, because I wanted to see who Claude would hire versus who I would hire. It was so different. Claude came up with a scoring system and said: based on these people's answers, how they use AI and everything else, these are the numbers one, two and three you should hire. I thought, okay, that's really interesting.
Then I brought everyone to the office. One thing I do is spend three to four hours with each person, because I want to see how you think. Each candidate met me, our team members and our investor. Every person agreed on the ranking, and it was not the same as Claude's. The person Claude ranked first wasn't even in our top five.
There's a company that raised a ridiculous amount of money, something like $30 million, and they're a one-person company. Their whole thing is, "We are a one-person company." On social media they're doing incredibly. But you look at the reviews, and every single person says, "I just want to speak to a human. There's 99% I can do myself, but for this 1%, I just want to speak to someone." You can't replace that.
That's what we're trying not to do at Moshi. We give you the data and the information. We do all the data engineering you wouldn't know how to do, because most people are not trained data scientists, data engineers and quants. But ultimately, you have to decide what trade you want to put on and how you want to manage it. We can give you the conviction and the data. That's what we're selling: you are going to be so much more confident in your trades, because you have blind spots. Why do you have blind spots? Because we watch way more data than is humanly possible. We'll tell you the connections, but ultimately you're the one who's going to put on the trade and sleep with that risk.
I think it's a really interesting time. Social media makes people see things a certain way, but the reality is that AI needs to be a tool for certain things, not a tool to replace everything.
So much of this resonates with me. I can't tell you how many times I've gone to a brainstorm or a workshop and people have shown up with the same ideas, because they all used the same system. It's like everyone shopped at the same stores and showed up in the same gray top and jeans.
It really is. And one funny thing about what you just said, everyone coming up with the same thing: we gave about 20 people a hiring assessment, and we told them to use whatever AI tools they wanted. We asked them to come up with a list of 10 target people we should speak to. Real people at real companies, from LinkedIn.
Bear in mind we interviewed people across the US, and LinkedIn has hundreds of millions of people. Can you believe that 90% came up with the same 10 companies?
I'm not putting anyone down, but we were hiring salespeople and a head of growth. If you're doing outreach to clients, think about that. Yes, I agree AI can help you. But think about what your difference is.
There's an idea I keep coming back to: the iPhone moment that Major League Hacking (MLH) CEO Swift has written about. Suddenly everyone thought they could be a photographer, but you still wouldn't let someone with an iPhone shoot your wedding, you know?
There's a lot we can do with AI, and there's a lot we don't have to do with it. It's simple: if I can get Claude to do something, then I will get Claude to do it. I'll get Claude to do the 90%. But that 10% is exactly where you don't want Claude.
You don't want AI there, because at the end of the day, if something goes wrong, you want to go talk to someone. You want to be able to say, "Hey, let's figure this out." You don't want to just talk to your chat model.
Top comments (1)
The build-versus-run gap is where many AI projects become operationally expensive. A useful operating dashboard would separate model quality from data freshness, ingestion failures, and manual intervention rate; otherwise a strong model can mask the reliability problem in the surrounding system.