Anekha Sokhal calls her startup, Moshi, "the program that I wish I had when I was trading." She spent most of her trading career in energy derivatives at JPMorgan. In an October 22, 2026 episode of the Latent Space podcast, she told host swyx how that experience shaped the product: which data it collects, how it decides what a trader should see, and why her team builds some of its own models instead of buying them.
From math to the energy desk
Sokhal studied MORSE at the University of Warwick, a degree that combines mathematics, operational research, statistics and economics (Warwick). People used to find it "so nerdy" that she hated telling them about it, she said. Now those subjects are, in her words, "what everyone calls AI." Her résumé dates the degree to 2011 to 2014, followed by a 2014 internship on JPMorgan's emerging-markets desk, work as a commodity analyst and, from 2017 to 2021, energy-derivatives trading.
She wanted a career that was fast-paced and full of data, she said. Commodities suited her because so much of the market turns on fundamentals, news and physical events, and she liked "building it up from zero." While trading, she said, she sensed a shift toward data, and she left JPMorgan to found a data company. A Fulbright scholarship brought her to the US. The Fulbright Commission lists her as the recipient of its Fulbright–Elsevier Data Analytics Award for Rice University's data-science master's program (Fulbright). Her interview came while gas prices were swinging wildly, she recalled, and because the interviewers' own electricity bills were rising, they could relate to her case for using data on the energy crisis.
In the US she went deep into machine-learning research. She is a co-author of GIQ, a benchmark that tests how well vision models reason about 3D shapes, in its ICLR 2026 camera-ready version (GIQ paper). She also said she worked in collaboration with NASA. Her résumé dates Moshi to June 2025.
What Moshi does
Moshi, Sokhal said, provides real-time insights for financial traders, starting with commodity, energy and macro markets. "We process more data than is humanly possible to tell you everything that matters to your portfolio right now," she said. When swyx asked for an order of magnitude, she offered her own illustration: in the four seconds it takes her to say a sentence, "Moshi has learned 300 new things."
The data design comes from how she traded. She split the market into the kinds of information a trader looks at when deciding which trade to put on and how to manage its risk. The first kind is structured public data. That includes price histories for futures contracts, with technical indicators such as moving averages that update constantly, and time series for fundamentals, the supply and demand figures that drive many commodity markets. Unstructured public data comes next: news, plus "alternative data" such as posts on X and audio from podcasts.
The third kind belongs to the user. "I can't come and tell you, hey, you should, you know, read this person because you have who you need," Sokhal said. Traders have the analysts and subscriptions they rely on and their firm's internal data. Some energy-trading companies Moshi speaks with, she said, have a person dedicated to counting the LNG ships passing through the Strait of Hormuz. The US Energy Information Administration estimates that about a fifth of global liquefied natural gas trade went through the strait in 2024, most of it from Qatar (EIA).
Moshi combines the public and private data, cleans and transforms it, and uses statistics and machine learning to tag it. On top of that it builds an ontology, a structured map of the things in a market and how they relate, and places everything in what Sokhal called the Moshi "n-dimensional graph." Users get the results two ways. Moshi pushes information, so that each time a trader opens it there is something new related to their portfolio. Traders can also pull answers by asking questions, and a language model returns, she said, answers "in 30 seconds, which would take you typically days."
Bloomberg, Claude and Netflix
swyx asked whether people call Moshi a new Bloomberg, "the 800 pound gorilla" of market data. Sokhal praised Bloomberg's chat as "a fantastic aspect of the platform." swyx explained it for listeners: traders transact billions of dollars in that chat using standardized phrasing, and his own entry into natural language processing, the field of getting computers to handle human language, was turning Bloomberg chat into a structured pricing sheet for option pricing. Bloomberg's product page describes language-processing tools that pull securities details and intent out of those conversations (Instant Bloomberg).
"I wouldn't say we're a new Bloomberg," Sokhal said. Admitting she did not know how controversial it was, she offered another description: "if Bloomberg and Claude and Netflix had a baby, it would be Moshi." Claude is Anthropic's AI assistant. Anthropic has been moving into finance itself, announcing in October 2025 a Claude for Excel preview and connectors to financial data sources such as LSEG market data (Anthropic).
The Netflix part is the recommendation system Moshi built in-house. A trader watching one market, Sokhal said, cares about the second- and third-order effects "which you wouldn't immediately know." Her example was an oil trader. If an energy-equities company does something that would indirectly affect oil, "you need to know that right now. You're not going to get that from like a news stream. You will get that from a connected graph." When two related pieces of information come out at the same moment, she said, the only things that can link them are a person or a system that imitates that person.
Netflix bases its suggestions on viewing history, ratings, the tastes of similar members, details about each title, and context such as the device and time of day (Netflix). As Sokhal describes Moshi, its recommendations instead follow links between market events and the positions a trader already holds.
Trust and noise
swyx brought skepticism from his own career. He traded currency options in London and later moved to the buy side, the funds that invest money, where his firm paid for alternative data. That included the Twitter "firehose," the full real-time stream of public posts that Twitter sold to companies with tools for filtering it (Twitter developer examples). He called it "stupidly expensive" and very noisy: "sometimes people do just talk random crap," and some try to manipulate markets. On the buy side, he added, the first question for any vendor is "are you looking at my data?"
On noise, Sokhal said users already have trusted voices they follow. In commodities those include suppliers and LNG facilities that post on X because it is the quickest way to give a status update, and users tell Moshi about them. Moshi filters out as much noise as it can. When it shows a claim, it attaches a confidence score along with "everything else happening around the conversation." Disagreement is useful too. If Reuters, Bloomberg and the Wall Street Journal report different numbers of ships on the move, "the fact that they don't agree is a point of data because is there an opportunity there? Because no one knows."
Private data, and building versus buying
Client data "sits everywhere," Sokhal said. For messy PDFs, Moshi built a vision-language model, a model that reads images and text together. Clients upload documents to their cloud version of Moshi, and the system breaks them down and rebuilds the contents in its database. Other data can only be reached by connecting to a client's systems, which she called "more of a conversation."
swyx asked why Moshi runs its own model when vendors such as Reducto offer document processing off the shelf. Sokhal called it "a question that a lot of companies are battling with right now" and gave two reasons. First, if someone hands Moshi their data, she does not want it sent to a third party without an agreement on how that party processes it and what it does with it. Vendors offer ways around that concern: Reducto's developer page advertises deployment inside a customer's own private cloud or on-premises, along with zero-retention policies (Reducto).
Her second reason was the team. A vendor is "a great solution" for a company without engineers, she said, but most of Moshi's team has worked with NASA in some capacity, "then they want to build the models." Not every company has that, she added, so the answer "really depends like where you are as a team."