You are the Team — Anekha Sokhal, Moshi

Anekha Sokhal, founder and CEO of Moshi, appeared on the Latent Space podcast with swyx recorded at AI Engineer New York, in what she described as her first time at the conference. Moshi builds real-time insights for financial traders, beginning with commodity, energy and macro markets, and Sokhal framed the product around a recommendation engine that surfaces what matters to a trader's book rather than another feed of headlines. She said the week after the conference filled with customer meetings, including follow-ups in Houston with energy firms.
Her path into the problem started in trading. Sokhal studied mathematics, operational research, statistics and economics at Warwick in England, then interned in investment banking before joining JPMorgan, where she worked on the emerging markets desk and later moved into commodities and energy derivatives. She left to found a data company, won a Fulbright scholarship to come to the US, and went on to machine learning research, including a publication at ICLR with Stability AI and work in collaboration with NASA, before starting Moshi.
On volume, Sokhal offered a rough order of magnitude: in the four seconds it takes to say a sentence, she said Moshi has learned 300 new things. The input mix is deliberately broad. It includes structured time series from futures markets and technical indicators such as moving averages, time series for supply and demand fundamentals, and unstructured material: news, X posts and audio from podcasts. Layered on top is private data belonging to the customer — the analysts a desk already trusts, subscriptions, internal feeds and niche operational sources, such as a company's own contact tracking the number of LNG ships moving through the Strait of Hormuz.
Moshi cleans and transforms that material, tags it, builds an ontology over it, and stores it in what Sokhal called an n-dimensional graph. Output is a piece of software that pushes new information to the user each time it is opened, plus a query interface she said returns in about 30 seconds answers that would otherwise take days. The recommendation layer is the part she emphasises: if a trader is watching one instrument, the second and third order effects may be what actually matter, and a connected graph can surface them where a chronological news stream would not. Her example is an energy equities company whose actions indirectly affect an oil trader's position.
Her description of the category is deliberately non-literal. She resists the "new Bloomberg" framing even while acknowledging Bloomberg's standing as the incumbent, singling out its chat as a genuinely strong feature and recalling that her own first exposure to natural language processing came from parsing Bloomberg chat into a structured pricing sheet for option pricing. Moshi, in her telling, is a different kind of software — market data plus an interface for working with it plus personalised recommendations — hence the Bloomberg, Claude and Netflix analogy.
The trust question comes up on two fronts. On social data, she said users often arrive with their own short list of trusted voices, particularly in commodities where suppliers and LNG facilities post status updates because it is the fastest channel available. Moshi attaches a confidence signal to such items, weighing a claim against what is happening around the conversation, and treats disagreement between outlets like Reuters, Bloomberg and the Wall Street Journal as itself informative rather than a defect to be smoothed away. On private data, the concern is whether the system is reading a customer's proprietary information. Integration takes at least two forms: a vision language model for messy uploaded PDFs, which are parsed and rebuilt in Moshi's database, and direct system connections handled case by case.
Sokhal also drew a distinction that product teams building internal AI tools may recognise. Anyone can now assemble a personal Claude-powered dashboard, but getting a colleague to adopt it turns the exercise from a modelling problem into a user interface problem, since the newcomer has to fold it into an existing workflow. Her thesis is that solving the interface and solving the recommendation problem are linked, and that platforms like Spotify and Instagram succeed partly by pushing content at people so naturally that they do not notice they are on the platform.
The conversation returned repeatedly to pressure on AI engineers. At the conference, Sokhal said many engineers described management telling them there is no longer a team — they are the team — and being asked to build alone a system that would normally occupy about ten people. Her own sequencing advice was to meet users before hiring, and she described using the conference itself to gather customer conversations, which then produced the post-event meeting schedule.
This suggests Moshi is competing less on model capability than on data plumbing and relevance. Assembling market time series, fundamentals, social posts, audio and a customer's own documents into a single graph is unglamorous integration work, and the perceived quality of the product will depend heavily on whether the recommendation layer is precise enough to justify attention. A likely trade-off is coverage against precision: a broad ontology of connected entities surfaces non-obvious second order links, but also risks noise, which is presumably why the confidence scoring and divergent-source handling exist. The 300-things-per-four-seconds figure is a throughput claim about ingestion, not a measure of accuracy, so it says little on its own about how good the recommendations are across markets with thin public data.
Why it matters: teams building AI products for information-heavy professional users are being pushed toward recommendation and delivery rather than raw summarisation, and Moshi is a concrete bet that the graph connecting entities, not the model, is where the differentiation sits. For product teams, the interface argument cuts against building bespoke internal dashboards nobody adopts, and points instead at pushing relevant material into existing workflows. For engineers told they are the whole team, the practical lesson implied by her account is to spend early time with users before scaling headcount.
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.