Anekha Sokhal on Moshi, a Recommendation Engine for Commodity Traders

The episode opens on Anekha Sokhal already mid-thought. She was supposed to leave New York the day of the recording. After AI Engineer, Thursday filled, then Friday, then Monday; energy people want her in Houston next. A week of meetings has grown out of the conference. The other thing she kept hearing in the halls is pressure: managers now tell AI engineers they do not need a team, because that one person is the team, and the system they are asked to build is the kind of thing ten people used to ship.

swyx cuts to the studio. This is his last AI Engineer podcast of the week, with Anekha from Moshi. He welcomes her; she thanks him for having her on.

He checks that he has her name right — Anekha, yes — and says her background is why he wanted the conversation. He used to trade; she is now a founder, and this is her first AI Engineer. He asks how she found the conference, what giving the talk was like, and what happened after.

Anekha thanks the AI Engineer team and says she is glad to be on the podcast. Everyone she speaks to is overwhelmed. Sharing her own path feels useful, because her undergraduate work was statistics and mathematics, which people now simply call AI, and because she is talking to another former trader.

swyx asks whether she did the financial and stochastic calculus. She did. The degree was MORSE at Warwick: mathematics, operational research, statistics and economics. At the time it sounded painfully nerdy and she hated explaining it. Now the same background reads as machine learning and AI, and the reaction has flipped.

She knew she wanted to trade. She does not know swyx’s story into the seat, but she wanted something fast and built on data. Once she understood what trading actually was, she could not imagine another career. A first internship in investment-banking trading led to JP Morgan, where she started.

She began on the emerging-markets desk, moved into commodities, and spent most of her career in energy derivatives. Because of the data background, she kept trying to break a market into parts. In commodities that means fundamentals, news and the physical world actually moving. She liked building the picture from zero, and while she was still trading she could already see the shift: data first, then the event.

That is when she left JP Morgan. She wanted to be a founder, and she wanted a data company. She got a Fulbright to come to the United States, which swyx interrupts as “just casually got a Fulbright.”

The origin story is drier than it sounds. She was in a yoga class in London. Her parents run a jewelry business; she had already side-hustled a computer-vision startup at university, virtual try-on for that trade. The question was how to become a founder who could talk about the technology, in the country where founding is easiest. This was 2020 and 2021. Fulbright had a data scholarship, one for the whole of the UK. That was the ticket.

The interview landed in the COVID gas-market spike. They asked how she saw the future. She talked about a global energy-market crisis and how data sat inside it. Interviewers could feel it in their electricity bills. That got her to the US.

Once here she went deep into machine-learning research, published at ICLR with Stability AI, and worked in collaboration with NASA. The interesting work, she says, is what led her to Moshi.

swyx asks what Moshi is. She is building the program she wished she had when she was trading. The company gives real-time insights to financial traders, starting with commodity and energy markets and with macro. It processes more data than is humanly possible, and it tells you everything that matters to your portfolio right now.

He wants types of data, and an order of magnitude. She starts with volume. It takes her four seconds to say a sentence; in that time Moshi has learned about 300 new things. That is the throughput. On types, she geeks out the way she used to trade. She breaks the market into every dataset a trader uses to put a trade on and to manage risk: structured and unstructured. There is the futures time series, and all the technical analysis around it — statistical indicators, moving averages — updating constantly.

Next is the time series for fundamentals, especially in commodity markets where demand and supply dominate. Then unstructured data. News is the obvious layer. Alternative data sits under it: X, audio from podcasts. swyx double-checks that she means Twitter; she does. That, she says, is the public availability of data.

The other layer is the user’s private data. When he was trading he had the analysts he actually trusted. She cannot walk in and tell a desk which voice to read. People have subscriptions they care about, and they have internal data. Some of the energy shops she talks to have their own person counting LNG ships through the Strait of Hormuz. That feed is personal to the user.

Moshi takes the public and private streams together, cleans them, transforms them, runs statistics and machine learning, tags them, and lays an ontology on top. The result goes into what she calls the Moshi n-dimensional graph. The product then pushes: every time you open it, you see something new related to your portfolio. You can also pull. An LLM, she says, will give you in thirty seconds answers that would otherwise take days.

swyx asks the branding question: do people call it a new Bloomberg, and does she hate that? He reaches for the 800-pound gorilla, then for whatever the British version of that phrase is. Bloomberg chat, he says, is a genuinely great part of the platform. Traders transact billions of dollars in that chat, with a standardized formula. His own entry into NLP was parsing Bloomberg chat into a structured pricing sheet for his logs, which is how they did option pricing.

She will not call Moshi a new Bloomberg. It is a different kind of software. The easiest way to say it, and she is not sure how controversial it is: if Bloomberg and Claude and Netflix had a baby, it would be Moshi. swyx catches on “Netflix” and asks if they have videos. They do not. They have a recommendation system.

The RecSys is internal. If you are a trader looking at one thing, you actually care about the second- and third-order effects you would not immediately know. Bring it back to an oil trader: an energy-equities company does something that will hit oil indirectly, and you need to know that now. A news stream will not give it to you. A connected graph will, because both facts are arriving at once. The only other way is a human who can hold both in their head. She wanted that when she was trading.

swyx notes the addictiveness of Netflix, which is a problem in that product. Anekha agrees it is interesting anyway. He thinks recommendation systems are one of the most important parts of AI. As a curator of AI Engineer he runs an LLM RecSys track every year. Engineers often do not show up, even when the speakers are the people who run RecSys for YouTube, Netflix, Instagram and TikTok. He cannot understand the empty room.

She thinks part of it is that those teams make the problem look easy. He cuts in: this is the most important system in the entire company. She agrees, and links it to what Spotify and Instagram do well.

The new default, she says, is that everyone can build their own Claude dashboard. Fine, if it is your brain in a box. Try to get a new person to put that box into their workflow and you are no longer solving a model problem. You are solving a UI problem: how they actually interact with it. Users do not always use the thing they asked for.

The platforms that work push data at you until you are hooked. You do not notice you are on them; you just want the next thing, then the thing after that. Inside Moshi they have been asking how UI relates to a recommendation system. The thesis is that if you solve one, you are also solving the other. If everyone is building software, the live question is how to build software people will actually use.

swyx fills in his own background, which he had forgotten to mention. He traded currency options in London, spent time in rates, and followed a similar map: start with EMEA, then maybe G10, at least that is the language in currencies. You ingest a huge amount of information. LLMs can probably process it better now. They did not have the tooling when he was doing this ten years ago.

He then moved to the buy side, where you start paying for alternative data. Balyasny was one of the first shops there to take the X firehose — the Twitter firehose — at something like $300,000 a month, for all tweets. It used to be free, he says. She is not sure it was ever free. It was free for researchers, he clarifies; a hedge fund had to buy it.

It is also very noisy. His eyes go up at the idea of trading on it. People talk nonsense, and people try to move markets by saying things that do not matter. They also had satellite data, credit-card data, the rest of the alternative stack. On the buy side his number-one question is: are you looking at my data? Probably not, he says, but how do you assure people, and how do you integrate with systems you do not even see?

Anekha hears two questions. The first is, with so much data, how do you know what is important? On X, that conversation comes up immediately, and it generalizes. In commodities especially, people have trusted voices they follow. Suppliers and LNG facilities are active there because it is the fastest way to put out a status update. Users tell Moshi who those people are. That is why they treat a given account as trusted: the desk already believes that person. swyx mutters that it is still sketchy.

They filter as hard as they can. When they show a claim, they attach a confidence score: someone is saying this, and here is everything else happening around the conversation. What users actually care about is divergence.

Take a simple count: how many ships are moving, how many planes are overhead. Reuters may say one thing, Bloomberg another, the Wall Street Journal a third. They may not agree. The fact that they do not agree is itself a data point. Is there an opportunity because nobody knows? That is how Moshi tries to present the information.

The second question is that people’s data sits everywhere, and integration is not one thing. The first kind is messy PDF data. Moshi has built a vision-language model to read it. Users upload to their cloud instance; the system breaks the documents down and builds them back up in Moshi’s database. Other data you simply have to connect to their systems, and that is a longer conversation.

swyx assumes they run the vision-language model themselves, and asks how important it is to own it versus using a vendor. There is no shortage of them; he names Reducto as a document-processing example. Is it important to have your own models, or do you just take whatever is off the shelf?

She says a lot of companies are fighting this right now. First: if someone gives you their data, you do not want to send it to a third party. There are systems that will do the work for you, but you need an agreement on how they process it and what they do with it. Second: if you do not have a team of engineers, a vendor is a great answer. If you do have a team — a fun fact, most of Moshi’s has in some capacity worked with NASA — those people want to build the models. Not everyone has that. It depends where you are as a team.

swyx turns to her talk. She apparently had a lot of inbound. This was AI Engineer’s first finance-focused session, and he had been genuinely unsure who they could get in the door, even though they pitched it hard to that crowd. Who did she actually talk to?

Back-to-back conversations, she says. What they expected is not quite what they got. In the past hour people have been messaging her on LinkedIn: they saw the talk, are you in Houston, can we meet; are you in New York, can we meet. Many of them are traders, at funds or in consulting.

AI engineers were the surprise. A lot of them feel overwhelmed by pressure from management. swyx asks, to do what? People now say they do not need a team: you are the team. So one person is supposed to build the whole system, a system that would typically have been done by ten people.

They come over and ask whether they need a swarm of agents, and what happens when something breaks. The more interesting complaint is: I built the system, people told me they needed it, it is in their email, nobody is clicking it, and they still complain they have the problem. What do I do? That was one cluster. The other was partnerships: you are doing this, we are doing that, let’s talk after.

Which is why she extended the New York trip — Thursday, Friday, Monday — and why Houston is now on the calendar. A week after the conference is already full. Her team kept pinging: another person signed up, another person signed up. She thinks it is super cool. swyx says oh my god.

He asks whether all these one-to-ones are actually interesting, or just inefficient. People need to ask things in private, but she could be building the product. She says they try to be helpful: take an initial meeting, understand the problem, then have a plan. They are already streamlining. Because of the inbound they are hosting their first webinar; they cannot always do one-to-ones.

He offers a buy-side pattern. When he was on that side, the banks set up the meetings. They came to little rooms like this, a round table, and he would meet five at once. A CFO doing investor relations would sit down with five funds; he would be one of the five.

He has not got this off the ground yet, but AI Engineer should do those meetings on site. People saw her talk and want to speak more; they should sign up and she should do five-on-one. The questions are probably the same. She lights up: they used to do this in finance. It is a good idea, including for a smaller reason — you often do not know what the person you are meeting looks like, so you end up taking pictures of the meeting point and calling them.

An assigned meeting room would fix that. This whole floor is already that, swyx says, but they had to pre-assign it to other speakers. He wanted dynamic allocation and does not yet know how to run it. The point is that she should not have to fly to Houston, and even if they are in New York she should not have to go to their office. Everyone comes here, gets the meetings done, and she is out.

She likes a 30-minute slot with a sign-up. He asks what else the conference could do better. She says, no joke, it is the best-organized conference she has been to: clear guidelines, a strong speaker team, breakout rooms, a speaker lounge that was never this good. He got a haircut. She did a facial and a massage. Cognition underwrote that, he says; they do want to spoil speakers, and this year went over and above. Organizing is one thing.

Business value and efficiency matter too. He wants the conference to be useful for both hiring and selling, and he does not know how to balance the two. What is more important to her right now, hiring or selling? They have just hired and they plan to hire, and they are always meeting people, but the priority is clients. Not hiring.

Hiring inbound is already continuous on LinkedIn. They put up one job, had to close it, it was open four days, and they had 2,500 applicants. That is its own problem. Meeting people who want to buy is different. Some people have the problem and do not want to buy; some have the problem and do. Finding that match is the work.

She has not found anyone who does that matching well. She likes his one-on-one idea. He is, he says, transferring what he used to have as a hedge-fund person into this market, which is AI and finance. They will try to do it next year, and try to be the definitive event. On the buying side he thinks he can help her better. On hiring he wants to help too: 2,500 is too many, and most of it is noise, so she needs a better filter. She has a strong opinion. He tells her to go ahead.

A lot of the volume is undercover: people shopping the role, or spying. You also get people who are legally not allowed to work. Visas, noise around visas. As a first-time founder she did not know those things when she started. She did not know the market. There is a lot of noise attached to that. He says they will try to get her higher signal. They want to be a high-signal conference.

He is glad she spent the time, and congratulates her on the traction. She thanks him, then asks how he moved into AI — whether he just decided to start. He left the hedge-fund world, moved into tech, and when the AI wave hit he realized he had to move very quickly to establish the field. He wrote the AI engineering post, then started this conference and the media around it, because at heart he is a media person.

He misses trading a little, but this is more fulfilling: he gets to meet real people instead of staring at a screen all day. She feels the same. They have to go. Thanks all around, and they are out.