The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi

Anekha Sokhal explains that she studied mathematics, operational research, statistics, and economics at Warwick—subjects that once sounded “so nerdy” to other people but are now commonly associated with AI and machine learning. She wanted a fast-paced, data-heavy career, so she pursued trading, first interning in investment banking and then joining JP Morgan. She began on the emerging-markets desk before moving into commodities and energy derivatives, where she analyzed markets by breaking them into fundamentals, news, physical events, and other measurable components. While trading, Anekha noticed a shift from simply having data to being able to use it systematically. She had already experimented with entrepreneurship through a computer-vision startup related to her parents’ jewelry business, including the idea of virtual try-on. Wanting to become a technology-focused founder, she won a Fulbright scholarship—the only data scholarship of its kind in the UK that year—and moved to the United States. During her graduate and research work, she published at ICLR with Stability AI and collaborated with NASA. Those experiences eventually led her to create Moshi, a company intended to build the software she wished she had while trading.

Anekha describes Moshi as a real-time insights platform for financial traders, initially focused on commodity, energy, and macro markets. Its goal is to process more information than a person can reasonably absorb and surface what matters to a trader’s portfolio. She gives a concrete scale comparison: in the roughly four seconds it takes her to say a sentence, Moshi has learned about 300 new things. The platform combines structured and unstructured information. Structured inputs include futures-market time series, technical indicators, moving averages, and fundamental supply-and-demand data. Unstructured inputs include news, alternative data, X/Twitter posts, podcasts, and users’ private information—such as preferred analysts, subscriptions, internal research, or proprietary observations about LNG ships moving through a particular strait. Moshi cleans and transforms these sources, applies statistics and machine learning, tags them in an ontology, and organizes them in an n-dimensional graph. The resulting software both pushes relevant updates to the user and lets the user ask questions. Anekha says its LLM can provide answers in about 30 seconds to questions that might otherwise take days to resolve.

When the host asks whether Moshi is a new Bloomberg, Anekha resists the comparison. She acknowledges that Bloomberg is the dominant platform for financial professionals and points to Bloomberg Chat as a powerful feature; she even recalls pricing Bloomberg Chat into a structured pricing sheet as part of her early work in natural-language processing. But she characterizes Moshi as a different kind of product—“if Bloomberg and Claude and Netflix had a baby.” The Netflix element is the recommendation engine. If an oil trader is looking at one company or data point, Moshi tries to identify the related second- and third-order developments that matter immediately. For example, an energy-equities company’s action might indirectly affect oil, even though a conventional news stream would not explicitly connect the two. Moshi’s graph is intended to make that relationship visible by linking information arriving at the same time. Anekha says the system imitates the useful part of a human trader’s ability to connect apparently separate facts, while also recommending information the trader would not have known to seek. The host agrees that recommendation systems are among the most important AI systems in companies, even though engineers often underappreciate them.

The host argues that recommendation systems are often overlooked because products such as YouTube, Netflix, Instagram, TikTok, and Spotify make them appear effortless. He says that, in practice, recommendations may be the most important systems in those companies. Anekha connects this to a broader problem with AI products: anyone can now build a personal dashboard, but getting another person to use it requires solving a user-interface and workflow problem. A dashboard may represent its creator’s own thinking, but it does not automatically become useful to someone else. By contrast, platforms that push relevant information to users create a habit: people encounter something useful without deliberately searching for it and then want more. Anekha says Moshi is trying to understand the relationship between the interface and the recommendation engine, because solving one should naturally help solve the other. The host adds that traders often ingest large amounts of information and that modern language models could process this material more effectively than the tools available when he traded currency options and rates. Both speakers conclude that building an AI system is not enough; it must deliver information in a form people will actually incorporate into their daily work.

The host recalls that, on the buy side, his firm paid for alternative data such as the X/Twitter firehose, which he describes as costing around $300,000 per month for hedge funds. He questions whether traders can rely on social posts because people may say random things or deliberately try to manipulate markets. He also asks whether Moshi can use a firm’s private data without needing to see or expose the underlying systems. Anekha says the central challenge is not merely collecting data but determining what is important. In commodity markets, users often have trusted individuals or organizations—suppliers, LNG facilities, or other industry participants—that use social media to provide rapid status updates. Moshi can incorporate those preferences and show a confidence score, along with the surrounding conversation. Conflicting reports from Reuters, Bloomberg, and The Wall Street Journal are not simply discarded: disagreement itself can indicate that the market does not yet know the truth and may therefore represent an opportunity or risk. For private documents, Moshi uses a vision-language model to interpret messy PDFs uploaded to a cloud version of the product, break them down, and rebuild the information into a database. Anekha says the company is cautious about sending customer data to third parties. Whether to build models internally or use vendors depends on the sensitivity of the data and whether a company has the engineering resources to maintain its own models.

Anekha says the conference generated more commercial interest than expected. Traders from funds, consultants, AI engineers, and potential partners approached her, and people messaged her on LinkedIn asking for meetings in New York or Houston. She extended her New York stay several times, filling the week after the conference with meetings and planning an energy-focused conversation in Houston. AI engineers were also interested because management increasingly tells them, in effect, “we don’t need a team, you are the team,” leaving one person responsible for systems that would traditionally require ten people. Some engineers asked whether they needed a swarm of agents and how to handle failures; another had built a system users claimed to need, but nobody clicked on it. The host and Anekha discuss making future conferences more efficient through scheduled group meetings, drawing on finance-style roundtables where one company meets several funds at once. Anekha says her priority is meeting customers rather than hiring. Moshi received 2,500 applications for one job posted for only four days, creating a filtering problem full of noise, including candidates applying while unauthorized to work or people effectively “shopping” or spying. The host says the conference should help create a higher-signal hiring filter. He then explains his own move from hedge funds into technology and AI, saying he misses trading somewhat but finds the conference and media work fulfilling because they let him meet people rather than stare at a screen all day.