The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi

Summary

• Moshi’s positioning is deliberately not “new Bloomberg” — founder Anekha Sokhal’s framing is “if Bloomberg and Claude and Netflix had a baby, it would be Moshi.” The Netflix part is literal: an internally built recommendation system that surfaces the “second- and third-order derivatives” of what a trader is watching — e.g., an energy-equities move that indirectly impacts oil — “right now,” which “you’re not gonna get from a news stream.”
• The throughput-and-breadth claim: “it takes me four seconds to say this sentence to you, and in the same amount of time, Moshi has learned 300 new things.” The stack spans futures time series and technicals, fundamentals supply/demand, unstructured news/X/podcast audio, and users’ private data (internal analysts, LNG ship counts through the Strait of Hormuz), fused into an “N-dimensional graph” that promises “answers in 30 seconds that would typically take you days.”
• Sokhal’s sharpest product thesis is that AI made it easy to build dashboards but not to get people to use them: “you’re no longer dealing with a problem, you’re dealing with a UI problem.” Anyone can now build “your brain in a dashboard,” but getting someone else to use it requires push-based hooks à la Spotify/Instagram — and Moshi’s internal bet is that “if you solve for one [UI], you’re naturally also solving for the other [the recsys].”
• swyx’s buy-side pushback is the episode’s best stress test: at Balyasny, he says they were among the first to use the X Firehose; he describes it as costing around $300,000 a month for all the tweets, and questions whether he’d trade on them — “sometimes people just talk random crap, and sometimes people try to manipulate markets.” Sokhal’s answer: commodity suppliers and LNG facilities post status updates on X as trusted sources, Moshi attaches confidence scores, and disagreement between Reuters/Bloomberg/WSJ “is a data point” — whether there’s an opportunity is unknown.
• On document AI, Moshi has built its own vision-language model for messy PDFs. Sokhal says data privacy is the first consideration when deciding whether to send client data to a third party, while noting that vendors can be useful with clear agreements and off-the-shelf solutions are a good option for teams without engineers. Her team wants to build models; most of it has, in some capacity, worked with NASA.
• Demand signals from her first AIE look real: she extended her New York stay three times, filled a week of post-conference meetings, and got requests to meet in Houston about energy. Hiring is a separate challenge: one job posting open four days drew 2,500 applicants. swyx called most of the applications noise; Sokhal pointed to undercover applicants and people not legally allowed to work, including visa issues. Her priority is meeting clients, not necessarily hiring.
• A recurring buyer persona: AI engineers overwhelmed by management’s new default — “We don’t need a team. You are the team.” They’re being asked to build alone systems “typically done by 10 people,” asking whether they need “a swarm of agents” and what to do when tools they’ve built go unclicked — a window into the problems they bring to Moshi.

Deep dive

1. From JPMorgan energy derivatives to a Fulbright in the U.S.

• Sokhal studied MORSE (mathematics, operational research, statistics, economics) at Warwick — “what everyone now calls AI.” After internships in investment banking and trading, including at JPMorgan, she started on emerging markets and moved to commodities and energy derivatives. Her trading style foreshadowed the product: breaking markets into their parts and building them up from fundamentals, news, and what was physically happening.
• A founder spark came from a computer-vision startup she side-hustled at university: with her parents in the jewelry business, she explored virtual try-on and wanted to become a founder who could talk about the technology. Around 2020 or 2021, she found Fulbright’s data scholarship, which she says was awarded to one person in the whole of the UK. In her interview, she discussed the energy-market crisis and volatile gas prices, which interviewers could relate to through their electricity bills. Fulbright brought her to the U.S., where she did machine-learning research, published at ICLR with Stability AI, and collaborated with NASA before Moshi.

2. What Moshi ingests — and the throughput claim

• The scale framing: “it takes me four seconds to say this sentence to you, and in the same amount of time, Moshi has learned 300 new things.” Sokhal calls it “the program I wish I’d had when I was trading”: real-time insights for commodity, energy, and macro traders.
• The taxonomy she geeks out on: futures time series with technical indicators and moving averages; fundamentals time series for supply and demand; unstructured news, X data, and podcast audio; plus the user’s private layer — trusted analysts, subscriptions, internal data, and examples such as tracking LNG ships through the Strait of Hormuz. Moshi cleans and transforms the data, applies statistics and machine learning, tags it under an ontology, and puts it into the “Moshi N-dimensional graph.”
• The output is both push and pull: each time a user opens Moshi, it surfaces something new related to their portfolio, and the LLM can provide answers in 30 seconds that would typically take days.

3. The Bloomberg-Claude-Netflix baby, and why RecSys matters

• Asked whether she’s building the “new Bloomberg,” Sokhal demurs — Bloomberg Chat is a “fantastic” feature — and calls Moshi a different type of software: “if Bloomberg and Claude and Netflix had a baby, it would be Moshi.”
• The Netflix piece is the recommendation system. A trader watching one thing may care about its “second- and third-order derivatives” — for example, an energy-equities move that indirectly affects oil. Sokhal says a news stream won’t surface that link; a connected graph can show it, and making the connection requires either a human or something mimicking a human, such as a recommendation system.
• swyx says he runs an LLM RecSys track at AIE every year featuring people behind recommendation systems at YouTube, Netflix, Instagram, and TikTok, but engineers often don’t appreciate how important RecSys is. Sokhal’s diagnosis: “they make it seem so easy.”
• Her adjacent thesis on the post-AI software glut: everyone can build “claw dashboards,” but getting someone else to incorporate one into their workflow is a UI problem. Platforms that push data can hook users; Moshi’s internal bet is that solving UI and solving the recommendation system are linked.

4. swyx’s buy-side interrogation: noisy tweets, private data, build-vs-buy

• The host’s credentials frame the pushback: swyx says he traded currency options in London, then rates, before moving to the buy side. At Balyasny, he says they were among the first to use the X Firehose; he describes it as “stupidly expensive,” around $300,000 a month for all the tweets. He says his eyes glaze over at the noisy data and questions whether he’d trade on it: people “talk random crap,” and some try to manipulate markets. His number-one buy-side question for Moshi: “Are you looking at my data?”
• Sokhal’s filter: commodity suppliers and LNG facilities are active on X because it’s a quick way to give status updates, and users have people in the industry whose voices they trust. Moshi attaches a confidence score and shows the surrounding conversation. Disagreement between Reuters, Bloomberg, and The Wall Street Journal is itself a data point; whether there’s an opportunity is unknown.
• For messy PDFs, Moshi has built a vision-language model to break down uploaded data and build it into its database. On build-vs-buy, Sokhal says the first consideration is not sending client data to a third party without an agreement on how it will be processed. She also says off-the-shelf systems are a good solution for teams without engineers; her team wants to build models, and most of it has worked with NASA in some capacity.

5. Post-talk inbound as a market read: overwhelmed AI engineers, 2,500 applicants

• After the talk, Sokhal was meant to leave New York that day but extended her stay for meetings on Thursday, Friday, and Monday. People also asked to meet in Houston about energy; she says the week after the conference filled up with meetings. The inbound included traders at funds or in consulting, AI engineers, and potential partners.
• The most telling buyer segment: AI engineers “overwhelmed by the pressure from management” because “people now say, ‘We don’t need a team. You are the team.’” They’re asked to build alone systems “typically done by 10 people,” and some ask whether they need “a swarm of agents” or what to do when a system is in users’ email but no one clicks it.
• One job posting open for four days drew 2,500 applicants. swyx says most of the applications are noise; Sokhal points to people applying undercover and people legally unable to work, including visa issues. She says the priority is meeting clients, not necessarily hiring, and that the team is hosting its first webinar because it can’t handle all the inbound one-to-one. swyx suggests on-site group meetings; Sokhal likes the idea of scheduled slots so people can meet without traveling to her or her team.