You are the Team — Anekha Sokhal, Moshi

What it claims

Moshi is a commercial product. Anekha Sokhal says it learns "300 new things" in the four seconds it takes to say a sentence. She also says its LLM interface answers in 30 seconds questions that would take a trader days. The page gives no evaluation behind either figure. Its main technical bet is a recommendation system that surfaces second- and third-order effects on a trader's book. Her example is an energy-equities event that indirectly moves oil. She argues a news stream misses this and a connected graph catches it. She also holds that UI and recommendation are one problem: "if you solve for one, you're naturally also solving for the other."

How it works

flowchart LR
  A[Futures time series + technicals] --> D[Clean, transform]
  B[Fundamentals: supply/demand] --> D
  C[News, X, podcast audio] --> D
  P[User PDFs via VLM] --> D
  S[Connected client systems] --> D
  D --> T[Stats/ML tagging + ontology]
  T --> G[n-dimensional graph]
  G --> U[Push: new items on open]
  G --> Q[Pull: LLM Q&A]

Private data enters by one of two routes. Users can upload messy PDFs to their cloud instance, where a self-built vision-language model breaks them down into the database. Other sources need a direct integration with the client's systems, which is handled case by case. For noisy social sources, Moshi filters first. It then attaches a confidence score that sets a claim against surrounding coverage. Disagreement between outlets such as Reuters, Bloomberg and the WSJ is kept as a signal in itself, not resolved.

Build versus buy

Sokhal gives two reasons for owning the document model rather than using a vendor such as Reducto. First, client data should not go to a third party whose processing terms must be checked. Second, Moshi has engineers who want to build models; she says most of the team has worked with NASA. She calls vendors the right choice for teams without that capacity.

Field notes from AI Engineer New York

AI engineers told Sokhal that management now says "you are the team." They are asked to build alone systems that would normally take 10 people. They asked whether to use agent swarms and what happens when those break. One engineer described a tool shipped to users' email that nobody clicks, although the same users still complain about the problem it solves. In the interview, swyx puts the X firehose at about $300,000 a month and calls it noisy and open to manipulation. Moshi closed one job posting after four days with 2,500 applicants.

Limits

This is an interview, not a technical write-up. It gives no accuracy, latency or recommendation-quality metrics. It does not give the size, architecture or training data of the vision-language model, or any pricing, API or deployment details. The linked "GIQ research paper" is listed in the show notes but never described.