The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi

“I'm so excited to just share, um, share my experience, and also I think we're living in such a fast-moving time, and everyone I speak to is, like, very overwhelmed, so it's really helpful to be able to kind of share everything that I've had, um, in my experience because I did my undergrad in statistics and math, which is now what everyone calls AI.”

Claim: Anekha Sokhal studied statistics and mathematics as an undergraduate and observes that people now commonly call that field AI.

“Did you do financial calculus? Like, like stochastic calculus? Yeah. I did that, and I-- So my undergraduate degree was mathematics, operational research- Ah... statistics, and economics, MORSE at Warwick. Okay.”

Claim: Anekha Sokhal earned an undergraduate degree in mathematics, operational research, statistics, and economics (MORSE) at Warwick.

“I don't know what, what your story to go into trading was, but I wanted something which was fast-paced and to do with data. Like, absolutely loved data. And it was-- When I read about what trading was, I could not imagine myself being in any other career, so I was just set on doing that.”

Claim: Anekha Sokhal chose trading because she wanted fast-paced work involving data and could not imagine herself in another career after learning about trading.

“So I got my first internship, um, in investment banking and trading, and then got an internship at JP Morgan, where I started, and I started on the emerging markets desk and then moved to commodities and, uh, did derivative trading there.”

Claim: Anekha Sokhal interned in investment banking and trading and later at JP Morgan, where she worked first on emerging markets and then in commodity derivatives trading.

“So most of my career has been in energy derivatives, and, uh, because I have that data background, I always looked at the market from... I wanted to break it down into all the different parts.”

Claim: Most of Anekha Sokhal's career has been in energy derivatives, and she used her data background to analyze markets by breaking them into component parts.

“So for example, in commodities, a lot is about the fundamentals and about the, uh, different news and the different, um, the physical, uh, happening. And so I was always fascinated by, like, building it up from zero.”

Claim: Anekha Sokhal was interested in constructing a commodity-market analysis from the ground up using fundamentals, news, and physical events.

“Actually, I wanted to be an entrepreneur and, um, and a founder, but I wanted to have a data company. So I got the Fulbright scholarship to come to the US because you guys have such an incredible venture space.”

Claim: Anekha Sokhal pursued a Fulbright scholarship to move to the United States and found a data company, viewing the United States as having an exceptional venture ecosystem.

“I side hustled a computer vision startup before my startup at university, and so I was like, "Oh, what are the problems in the jewelry space? Oh, like, virtual try-on." So that led to me wanting to be a founder, and I thought, "How could I be a founder in the best country to be a founder in?" And I thought, "What are... I, I wanna be a founder who can talk about the technology."”

Claim: Anekha Sokhal's university computer-vision side project explored virtual try-on for her parents' jewelry business and contributed to her decision to become a founder.

“So this was back in, like, twenty twenty, twenty twenty-one. And I found Fulbright had a da- data scholarship, and they gave one in the whole of the, the UK. Uh-huh.”

Claim: In 2020–2021, Anekha Sokhal found that Fulbright offered a data scholarship with one award available across the United Kingdom.

“Mm-hmm. So there was, there was... Like, gas markets had ig- insane volatility. And so when I did my Fulbright interview, they asked me, "How do you f- Like, how do you see the future?" And I talked about the, like, global issue of everyone is having the energy market crisis and how data relates to that. And everyone could relate to what I was talking to because, like, gas prices were crazy.”

Claim: During her Fulbright interview amid the 2020 energy-market crisis, Anekha Sokhal discussed how data related to the crisis, when gas prices were unusually high.

“Their electricity bills were going crazy. Um, and so that led me to the US. Came here, was deep in machine learning research, published at ICLR with Stability AI, worked in collaboration with NASA, did a lot of, a lot of super interesting work, which then led me to Moshi. And what is Moshi?”

Claim: After moving to the United States, Anekha Sokhal conducted machine-learning research, published at ICLR with Stability AI, collaborated with NASA, and subsequently joined Moshi.

“So Moshi, we are building, um, the program that I wish I had when I was trading. So what we do is we provide real-time insights for financial traders, starting with commodity and energy markets and macro markets.”

Claim: Moshi provides real-time insights for financial traders, starting with commodity, energy, and macro markets.

“So what we do is we process more data than is humanly possible to tell you everything that matters to your portfolio right now. What kind of data would you be processing? Or like give us an- Yeah...”

Claim: Moshi processes more data than a human can handle to identify what matters to a trader's portfolio in real time.

“So a fun way I explain it to people is it takes me four seconds to say this sentence to you, and in the same amount of time, Moshi has learnt 300 new things. So that's the type of volume of data that we process.”

Claim: Moshi learns 300 new things in approximately four seconds, according to Anekha Sokhal.

“I would break the market up into all the different types of m- data, which a trader looks at to decide what trade to put on and how to manage the risk.”

Claim: Moshi organizes market data into categories relevant to traders' decisions about trades and risk management.

“So what I do is I take the market and I say, well, there's structured data and unstructured data. There's, there's the time series of the futures market, and there's all of the technical analysis around it.”

Claim: Moshi analyzes structured and unstructured market data, including futures time series and technical-analysis information.

“So we're looking at all of this, like statistical indicated, all of the different technical indicators, all of the moving averages. That's one data set. And obviously, that is updating constantly.”

Claim: Moshi analyzes statistical indicators and moving averages that update continuously.

“The next data set we look at is the time series for fundamentals markets. So especially in, um, some commodity markets, there's a lot of demand and supply. And then we look at the unstructured data.”

Claim: Moshi analyzes time-series data on commodity-market fundamentals, including supply and demand.

“So the news is the obvious one. Mm-hmm. But then there's also all this alternative data that exists. For example, there's X data, there's audio data on podcasts. Ex-Twitter. Ex-Twitter, yes. You read Twitter for... Okay.”

Claim: Moshi analyzes unstructured information from news, X, and podcast audio.

“So understanding all of these different types of data, and that I would say is the, like, public availability of data, and there's this whole other type of data, which is the user's private data. Mm-hmm. So if, I'm sure you had that when you were trading, which is you have the analysts you care about. I can't come and tell you, "Hey, you should, you know, read this person," 'cause you have who you need. And then you're gonna have your, um, subscriptions that you care about. You'll have your internal data, right? Like some of the companies we speak to, if they are, uh, trading markets, trading energy markets, they will have their person for the amount of LNG ships going through the Strait of Hormuz. And so there's all this data which is personal to the user.”

Claim: Moshi combines public data with user-specific private data, such as selected analysts, subscriptions, and internal information.

“So we take those two data points together, do all of the, like, data cleaning, data transformation.”

Claim: Moshi cleans and transforms the public and user-specific data it collects.

“We then look at the statistics and machine learning, tag it, kind of create our whole ontology on top of it, and then we put it into the Moshi, like n-dimensional graph, and then we pro- produce the output, which is a piece of software which pushes data to you.”

Claim: Moshi uses statistics and machine learning to tag and structure data in an ontology and n-dimensional graph, then delivers it through software that pushes information to users.

“So every time you open Moshi, you see something new related to your portfolio, and you can pull data. So LLM, I'm gonna give you answers in 30 seconds, which would take you typically days.”

Claim: Moshi surfaces new portfolio-related information when opened and can answer some questions in 30 seconds that would typically take days to answer.

“Um, I wouldn't say we're a new Bloomberg. I'd say we're a different type of software which is not, I think this is the easiest way to say it, and I don't know how controversial this is, but if Bloomberg and Claude and Netflix had a baby- Netflix?... it would be Moshi. You have videos? Netflix.”

Claim: Anekha Sokhal characterizes Moshi as a different kind of software rather than a new Bloomberg, comparing it to a combination of Bloomberg, Claude, and Netflix.

“We actually have a recommendation system- Ah... which we've internally built, which is if you're, if you're a trader and you're looking at one thing, um, you actually care about the second and third order derivatives right now, which you wouldn't immediately know.”

Claim: Moshi's internally built recommendation system identifies second- and third-order market effects relevant to the item a trader is viewing.

“And so we've built a whole recommendation system where a simple thing, if you're, um, let's bring it back to, like, an oil trader, and there's an energy equities company that's doing something which would di- indirectly impact oil, you need to know that right now.”

Claim: Moshi's recommendation system can surface an energy-equities company's activity to an oil trader when that activity may indirectly affect oil.

“Uh, you're not gonna get that from, like, a news stream. Mm-hmm. You will get that from a connected graph which is showing that to you.”

Claim: Moshi's connected graph can reveal indirect market connections that a conventional news stream would not provide to a trader.

“Because if those two information points are coming out right now, the only way you can do it is either a human, which is being able to, like, connect those, or by a mimicking human, sort of like a recommendation system which can do that.”

Claim: Connecting separate but relevant information points requires human reasoning or a human-like recommendation system, according to Anekha Sokhal.

“That's what we try and do because I, like, wish I had that when I was trading. Yeah. Yeah. Well, I, I, you know, there's a whole addictiveness element of Netflix which is bad. [laughs] There is.”

Claim: Anekha Sokhal wishes she had access to Moshi's capabilities during her trading career.

“I think, like, REX/X is one of the most important parts of AI, and as a curator of AI years, I actually have every year an LLM REX/X track, and engineers often don't appreciate how important it is.”

Claim: The interviewer believes recommendation systems are among AI's most important components, yet engineers often underestimate their importance.

“Like, it's not very well attended, uh, even though some of the people powering these things, I mean, they run REX/X for YouTube, they run Netflix, uh, for Netflix, for, uh, Instagram and TikTok.”

Claim: Recommendation systems are used in products at YouTube, Netflix, Instagram, and TikTok, according to the interviewer.

“And then, like, nobody shows up, and I'm like, "What do you mean?" Like, [laughs] "You're not interested in REX/X?" You know, I, I find it really interesting. I think part of it is it, they make it seem so easy. Uh-huh.”

Claim: The interviewer suggests that recommendation systems' apparent simplicity may contribute to people underestimating their importance.

“And it's like, oh, this is so, like, why- This is the most important f- system in the entire, like, you know, company is this thing. Exactly.”

Claim: The interviewer argues that recommendation systems may be among the most important systems in a company.

“And I'd say there's arguably something linked to it, which I think some companies do really well, like Spotify, Instagram, which is- One thing that people have come out with AI is everyone now can build their own Claude dashboards, right? It's like, "Hey, I built this myself." Great. If you want to get someone else to use it, that's your brain in a dashboard or your brain in something. Try and get someone new to incorporate that into their workflow.”

Claim: A self-built AI dashboard will not gain another user's adoption unless it is incorporated into that user's workflow, according to Anekha Sokhal.

“You're then not lo- no longer dealing with a problem, you're dealing with a UI problem, which is how did they interact with it? 'Cause sometimes users don't always use that.”

Claim: After a product is built, adoption can become a user-interface problem: users may not interact with it or use it as intended.

“One thing that these platforms do really well is they push data to you, which, like, hooks you, so you naturally don't even realize you're on the platform, and you're like, "Oh, wait, I want that.”

Claim: Platforms that proactively push data to users can draw users in and prompt them to seek more information.

“No, no, I actually want that." And so it's... We've had-- We've been trying to internally as a company understand, how does UI relate to a recommendation system? Because our thesis is if you solve for one, you're naturally also solving for the other. Mm.”

Claim: Moshi's thesis is that solving the user-interface problem and solving the recommendation-system problem are naturally linked, such that solving one also addresses the other.

“And I think that's a big thing right now, which is if everyone's building software, how do you get, get it to actually be software that people will use? Yeah. It's, it's a bit of an interesting [chuckles] take. Yeah.”

Claim: A key challenge in software development is ensuring that people actually use the software being built.

“And, um, i-it's, uh, you, you ingest a, a lot of information, and really probably LMS can process it, uh, better. Um, so we didn't have the tooling back when I did this 10 years ago. Mm-hmm.”

Claim: The interviewer believes language models can probably process the large volume of information traders ingest more effectively than traders can.

“And, um, i-it's, uh, you, you ingest a, a lot of information, and really probably LMS can process it, uh, better. Um, so we didn't have the tooling back when I did this 10 years ago. Mm-hmm.”

Claim: The interviewer says traders lacked the relevant information-processing tools ten years earlier.

“But, uh, now I, I see that, and then, then I move to the buy side, where you start paying for alternative data, exactly like, like you, like you did. Uh, we were actually one of the very first few at, at Balyasny to do the X Firehose or Twitter Firehose. Yeah. And that, that's like something like a stupidly expensive, like, $300,000 a month, like, uh, data thing that you have to pay for to get all tweets, um, and- Used to be free... and you sort of parse them.”

Claim: The interviewer says his buy-side firm paid for alternative data such as Twitter Firehose, which he recalls costing approximately $300,000 per month.

“And it's very noisy, and, and some... And so then my, my eyes goes up because I don't know if I would trade on those things because- Mm... sometimes people do just talk random crap, and sometimes people try to manipulate markets by- Yeah... saying stuff, and it just doesn't matter that much.”

Claim: The interviewer considers Twitter data noisy because users may post irrelevant or market-manipulative content, making it of limited trading value.

“But we, we would also have, um, alternative asset, uh, data like satellite information or credit card information or whatever, right? And so my, on the buy side, my number one question is, "Are you looking at my data?"”

Claim: Buy-side firms may use alternative data sources such as satellite imagery and credit-card information, according to the interviewer.

“One thing we found with people is they really will have their trusted people who they'll follow of these voices in industry.”

Claim: Moshi has found that users, especially in commodity markets, follow trusted industry voices as information sources.

“Especially in commodities, they'll be the suppliers of commodities, they'll be the LNG, uh, the, the... all of these facilities who will be active on there because it's the quickest way for them to also give a status update. Wow. So it's interesting because we found that a lot of our users will tell us about this.”

Claim: Commodity suppliers and facilities use industry channels as a fast way to provide status updates.

“And so it's, it's one of those things where we have found a way to, uh, try and filter out as much as possible, but when we are showing it to someone, we give a, like, confidence score of, okay, someone is saying this one thing, but this is everything else happening around the conversation.”

Claim: Moshi filters information and provides a confidence score with the surrounding conversation to help users assess a claim.

“Because what's interesting to people is if there are divergent sources. Like, a, an example would be, um, how many ships are going up, how many planes are going above us right now. If Reuters is saying something or Bloomberg is saying something and Wall Street Journal is saying something, they may not all agree. Mm-hmm. But the fact that they don't agree is a point of data because is there an opportunity there? Because no one knows. So that's how we try and give people information. And then the second question is, you know, people's data sits everywhere, and I think there's different types of integration. The first one is, what about, like, their messy PDF data?”

Claim: Disagreement among sources such as Reuters, Bloomberg, and the Wall Street Journal can itself be useful information because it may reveal an opportunity amid uncertainty.

“Um, that data is what we've built is, like, a vision language model to understand it.”

Claim: Moshi uses a vision-language model to interpret users' unstructured PDF data.

“So we'll take the data that they can upload to their cloud version of Moshi, and then we actually go and we break it down and then build it up in our database.”

Claim: Moshi processes data uploaded to its cloud version by decomposing it and incorporating it into Moshi's database.

“And there are certain types of data where you just have to connect to their systems. Um, and then that's, like, more of a conversation. Is it important to, um...”

Claim: Moshi must connect directly to some customers' systems to access certain types of data.

“The first question for us is we-- If someone's giving us their, their data, we don't wanna be sending it to a third party.”

Claim: Moshi does not want to send customer-provided data to third parties.

“Because there are systems out there which can do these things for you, um, but you make, you wanna make sure that you've got that agreement on how are they processing that data, what are they doing with it.”

Claim: Companies should have clear agreements with third-party data processors about how customer data is processed and used.

“Um, that's the first question. The second one is, if you don't have a team of engineers, then that is a great solution.”

Claim: Using third-party systems can be a good option for companies without an engineering team.

“But if you do have a team of engineers like us, um, a fun fact is most of the team has, in some capacity, worked with NASA. Um, then they want to build the models- Mm-hmm But not everyone has that. And so, you know, I think it really depends, like, where you are as a team. Yeah. Got it. Um, okay, and then coming into your talk, you, uh, apparently had a lot of inbound from it. Yeah.”

Claim: The choice between building models in-house and using external solutions depends on a team's capabilities and circumstances.

“But if you do have a team of engineers like us, um, a fun fact is most of the team has, in some capacity, worked with NASA. Um, then they want to build the models- Mm-hmm But not everyone has that.”

Claim: Companies with engineering teams may choose to build their own models, but many companies lack that capability.

“Or, "Hey, are you in New York? We- can we get a meeting?" The types of people we've had, um, a lot of traders, either at funds, um, consulting, and then AI engineers was an interesting one because a lot of them are feeling, like, overwhelmed with the pressure from management. To do what? Because people now say, "We don't need a team, you are the team." And so it's like, hey, I need to now build this whole system myself, but that system would be typically done by 10 people.”

Claim: Anekha Sokhal reports that many AI engineers she met felt overwhelmed by management expecting one person to build a system that would typically require ten people.

“So a lot of people have talked to us and they're like, "Hey, I, I need to build a system, but, like, you know, do I need to have a swarm of agents? But then, like, what if something breaks?"”

Claim: People have asked Anekha Sokhal whether building a system requires a swarm of agents and how to handle failures if something breaks.

“Something interesting is, "Hey, I built this system. People told me they needed it. It's in their email. No one's, no one's clicking it. But they complain that they have this problem. Like, what do I do?"”

Claim: Some builders report that users say they need a system but do not use it when it is delivered, despite continuing to complain about the underlying problem.

“So we have got things that we're doing now to stream that because, uh, like we're hosting our first webinar. Because of the amount of inbound that we have, we just can't always- Yeah, exactly... do these one-to-ones.”

Claim: Moshi is developing scalable outreach, including webinars, because it cannot accommodate all inbound interest through individual meetings.

“Uh, I mean, a- anything else that, like, we, I guess we could do better or, like, other experiences from the conference that you're- I think this has been-... on your mind... the best organized conference I've been to.”

Claim: Anekha Sokhal considers this the best-organized conference she has attended.

“So in terms of we've just hired and we plan to hire, um, we're always meeting people, but the priority right now is to meet clients. Yeah. That's a priority, not necessarily hiring.”

Claim: Moshi's current priority is meeting clients, not hiring.

“Um, because I think what we have with hiring is we get continuous inbound from LinkedIn. Really? Yeah. I had... We had one job up, we had to close it. It was open for four days. We had 2,500 applicants. Okay.”

Claim: Moshi received 2,500 applicants for a LinkedIn job posting that was open for four days, and the company receives continuous hiring interest through LinkedIn.

“Because I think there's people who have the problem, but they don't wanna buy, and there's people who have the problem and want to buy, and then you who are trying to find that match, that is- I haven't found any person who's done that- Yeah...”

Claim: Potential customers with a problem differ in whether they are willing to buy a solution, making it difficult to identify the right customer match.

“I think I, I think I can definitely help you better there. Uh, on the hiring side, I also want to help you, right? Okay. Yeah. Like 2,500 is too much and most of it's noise, and so you need a better- Oh... filter.”

Claim: The interviewer considers 2,500 job applicants excessive and mostly low-signal, and says a better screening filter is needed.

“Uh- I have a strong opinion. [laughs] Okay, go ahead. The amount of people who are also doing it in a way where it's, like, undercover, not, it's- They're shopping? They're spying? You have that, but you also have, it's, they're, like, legally not allowed to work, so there's so much noise in that. Oh. It's like- Visas... visas- Okay... and there's so much noise around that as well.”

Claim: Hiring applicant pools may contain people covertly shopping or spying and people legally unable to work, which adds noise to the process.

“And so when I was a first time f- like, I am a first-time founder. W- When I first started, you don't know those things. Yeah. You don't know the, like, market, right? And there's a lot of noise related to that. Yeah.”

Claim: As a first-time founder, Anekha Sokhal initially lacked knowledge of the market and the sources of noise in it.