AivexaNewsSearch
AI news for builders and product teamsChecked every hour

You are the Team — Anekha Sokhal, Moshi

Collected Oct 22, 2026

Anekha Sokhal, founder and CEO of Moshi, joined swyx on the Latent Space podcast during AI Engineer New York to talk about her route from energy derivatives trading at JPMorgan into building an AI product for traders. The episode was billed as swyx's final AIE podcast of the conference. Alongside the interview, the listing notes that tickets and applications are now live for AIE CODE, described as the last flagship AI Engineer event of the year.

Sokhal's background runs through mathematics rather than finance first. She studied mathematics, operational research, statistics and economics at Warwick in England, then interned in investment banking and trading before joining JPMorgan, where she started on the emerging markets desk and later moved to commodities and derivative trading. Most of her career, she says, has been in energy derivatives. A Fulbright data scholarship, which she describes as awarded once across the UK, brought her to the United States; she cites work published at ICLR with Stability AI and a collaboration with NASA before founding Moshi.

Moshi's stated product is real-time insight for financial traders, initially aimed at commodity, energy and macro markets. The pitch is that it processes far more information than a person could read and surfaces what matters to a specific portfolio. To convey scale, Sokhal offers an analogy: in the four seconds it takes to speak one sentence aloud, Moshi has learned 300 new things.

The data pipeline she describes has four rough layers. First, structured market data: futures time series plus the technical layer around them, including statistical indicators and moving averages, which update continuously. Second, fundamental time series, which matter heavily in commodities where supply and demand drive prices. Third, unstructured public data: news, posts on X and audio from podcasts. Fourth, and perhaps most distinctive, each user's private data — the analysts they already trust, their subscriptions and internal sources. Her example is a firm that keeps its own person tracking how many LNG ships are passing through the Strait of Hormuz. Telling a trader which analyst to read would be pointless, she argues, because they already have their own.

On top of that, Moshi performs cleaning and transformation, applies statistics and machine learning to tag items, builds an ontology, and stores everything in what Sokhal calls an n-dimensional graph. The output is software that pushes new material related to the user's portfolio each time it opens, and also answers queries — responses she says take about 30 seconds where the equivalent manual work would take days.

Asked whether Moshi is a "new Bloomberg," Sokhal declines the framing. She praises Bloomberg's chat as a genuinely strong part of that platform, noting that traders transact billions of dollars through it. Her own entry into natural language processing, she says, was parsing Bloomberg chat into a structured pricing sheet used for option pricing. She prefers a different description: what you would get if Bloomberg, Claude and Netflix had a baby — market data, a conversational interface for working with it, and recommendations.

The Netflix component is a recommendation system built in-house. Its premise is second- and third-order effects: a trader focused on one instrument should also know about adjacent signals they would not think to check. Her illustration is an oil trader who needs to know immediately when an energy equities company does something that indirectly bears on oil. A chronological news feed will not connect those dots; a connected graph, or a system imitating the human who would make the connection, can.

Sokhal is candid about noise in social data. Filtering is imperfect, so when Moshi surfaces a claim it attaches a confidence score and shows what else is happening around the conversation. Disagreement is itself informative: if Reuters, Bloomberg and the Wall Street Journal report a figure such as the number of ships or planes in a region differently, the divergence is a data point that may point to an opportunity, precisely because nobody knows the true number.

On integration with customer systems, she splits the problem in two. Messy PDF data uploaded to the cloud version of Moshi is handled by a vision language model that breaks the documents down and rebuilds them in Moshi's database. Other data classes require direct connection to customer systems, which she frames as a conversation rather than a product feature.

The interview also takes up build-versus-buy decisions for models and document extraction, and the pressure AI engineers reported feeling at the conference: management telling them there is no team, that they are the team, and expecting one person to assemble a system that would normally occupy about ten. Reflecting on her own conference experience, Sokhal says her New York stay kept extending as customer meetings filled the days after the event, including inbound interest from energy firms suggesting talks in Houston. Her takeaway on company-building is that meeting users comes before hiring.

swyx, for his part, draws on his own trading past — currency options in London and rates, then the buy side — recalling early work with the full X firehose at Balyasny, data he describes as extremely expensive (roughly $300,000 a month) and very noisy, and warning that some participants post nonsense or try to move markets. His recurring question as a buyer was whether a vendor was looking at his data, a concern that maps directly onto Moshi's mixing of public and private sources.

Why it matters: Moshi is a concrete case of applying recommendation systems — a discipline swyx notes gets thin attendance at AI engineering conferences despite running the world's biggest consumer platforms — to a domain where the value is in connecting correlated signals rather than summarizing documents. For developers building vertical AI products, the interesting claims are architectural: an ontology plus graph layer over heterogeneous public and private feeds, confidence scoring for noisy sources, and a vision language model for PDFs. The open question, which the product's success will answer rather than the interview, is whether a recommendation-first interface displaces or merely wraps the incumbent terminal workflows traders already trust.

Read at Latent Space

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

How Moshi turns market data into actionable insights, and why AI tools have to fit the trader’s workflow.