You are the Team — Anekha Sokhal, Moshi

What it claims

Anekha Sokhal, a former JPMorgan energy derivatives trader, describes Moshi as a recommendation engine for commodity, energy and macro traders. She frames it as "Bloomberg, Claude and Netflix had a baby": market data, an LLM interface for querying it, and a recommender that pushes items tied to a trader's book. Her stated throughput is that Moshi "has learned 300 new things" in the four seconds it takes to say a sentence. She also says LLM queries return in about 30 seconds answers that would typically take a trader days. Neither figure comes with a definition or measurement method.

How it works

Inputs fall into two groups. Public data covers futures time series with technical indicators and moving averages, fundamentals (supply and demand) time series, news, X posts and podcast audio. Private data is each user's own: chosen analysts, subscriptions and internal feeds, such as a client's own count of LNG ships through the Strait of Hormuz.

Both groups go through cleaning and transformation, then statistical and ML tagging into an in-house ontology. The result is loaded into what she calls an "n-dimensional graph." The recommender walks this graph to surface second- and third-order effects. Her example is an energy-equities event that indirectly moves oil, which a news stream alone would not flag for an oil trader.

Two design choices stand out:

• Confidence scoring. Social posts from sources users trust, such as LNG facility operators on X, are shown with a confidence score and with the surrounding conversation. When sources diverge (Reuters, Bloomberg and the Wall Street Journal disagreeing on a count), the divergence itself is presented as signal.
• Document extraction. Messy PDFs that users upload to their cloud instance are parsed by an in-house vision-language model. Some other data needs direct system integration, which is negotiated per client.

Build versus buy

Sokhal gives two reasons for building the VLM rather than using a vendor such as Reducto. The first is that client data should not be sent to a third party whose processing terms are unclear. The second is that her team, most of whom have worked with NASA, has the capacity to build models. She says buying is the right choice for teams without engineers to spare.

Field observations

AI engineers she met at AI Engineer New York reported pressure from management to build "10-person" systems alone ("you are the team"). A recurring failure was internal tools delivered by email that users asked for but never clicked. Her thesis is that the UI and the recommender must be solved together, with data pushed to users rather than placed in dashboards. One Moshi job posting drew 2,500 applicants in four days and was closed.

Limits

This is a conversational interview with an automated transcript. It gives no model architecture, sizes, evaluation results, pricing or customer metrics. It names a "GIQ research paper" without describing it.