The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi

What it claims

Moshi is a commercial product, not an open model or repository. Anekha Sokhal describes it as "Bloomberg, Claude and Netflix" combined: market data, an LLM interface over it, and a recommendation layer. It starts with commodity, energy and macro markets. Her throughput claim is that in the four seconds it takes to say a sentence, Moshi "has learned 300 new things." She says LLM queries answer in 30 seconds what would typically take a trader days. No benchmark, accuracy figure or evaluation backs either number.

How it works

The interview names the input classes:

• Market time series: futures prices with technical indicators and moving averages.
• Fundamentals: supply and demand time series.
• Public unstructured data: news, X posts and podcast audio.
• Private user data: a trader's chosen analysts, subscriptions and internal datasets. One example is a client's own tracker of LNG ships through the Strait of Hormuz.

flowchart LR
  A[Public structured + unstructured] --> C[Clean, transform, ML tagging]
  B[User private data, PDFs via in-house VLM] --> C
  C --> D[Ontology + n-dimensional graph]
  D --> E[Push: new items for the book]
  D --> F[Pull: LLM Q&A]

The recommender targets second- and third-order effects. An oil trader would be shown news about an energy equities company that indirectly affects oil, linked through the graph rather than a news stream. Users often name trusted X accounts, such as LNG facilities posting status updates. Moshi shows these with a confidence score set against the surrounding conversation. When Reuters, Bloomberg and the Wall Street Journal disagree, Sokhal treats the disagreement as a signal in its own right. Her thesis is that the UI and the recommender are one problem: pushing items into the workflow beats a dashboard nobody opens.

Build versus buy

Moshi runs its own vision-language model for clients' messy PDFs. It does not use a vendor such as Reducto, which swyx raises as an option. Her first reason is that client data should not go to a third party without clear terms on how it is processed. Her second is that the team can build it, since most members have worked with NASA. She says teams without engineers should buy. PDFs come in through uploads to a client's cloud instance. Other client systems need custom integration.

Limits

The episode is an interview with no architecture details, model names, sizes, latency or evaluation results. swyx adds that the X firehose cost about $300,000 a month at Balyasny and calls it noisy and open to manipulation. The episode does not say how Moshi's filtering handles that.