Key Takeaways

  • Commodities trading desks make decisions by pairing public pricing feeds with strictly private operational data, such as tracking LNG tankers through the Strait of Hormuz.
  • Moshi ingests structured futures time series alongside alternative streams from podcasts, news, and X, learning 300 new data points every 4 seconds.
  • In-house vision-language models process complex PDF reports locally to extract domain entities without sending private intelligence to third-party model APIs.
  • Transformed entities feed into an n-dimensional knowledge graph that links distant asset movements, resolving in 30 seconds research that previously took days.
  • Trading teams organize this data flow using Moshi's Financial Intelligence Ingestion and Delivery Pipeline.

The Moshi Financial Intelligence Ingestion and Delivery Pipeline

Trading commodities requires absorbing volatile market signals without leaking proprietary positions. Anekha Sokhal explains the velocity required: “It takes me 4 seconds to say the sentence to you and in the same amount of time, Moshi has learned 300 new things.” To keep up, Sokhal structures market analysis across public feeds and private intelligence:

Step 1: Public Market Data Ingestion

Ingest structured time series of the futures market (technical indicators, moving averages) alongside fundamental supply/demand metrics and unstructured alternative data (news, X/Twitter commentary, podcasts).

Step 2: Private Proprietary Data Ingestion

Isolate and ingest the user's private data, including trusted analyst lists, paid subscriptions, and internal operational data (e.g., proprietary tracking of LNG ships through the Strait of Hormuz) without exporting it to third-party APIs.

Step 3: Cleaning, Normalization, and Ontology Tagging

Parse messy unstructured documents (like complex PDFs via in-house vision-language models), clean and transform data points, run statistical checks, and apply machine learning tags to map entities onto a unified domain ontology.

Step 4: Knowledge Graph Synthesis

Deposit transformed entities into an n-dimensional connected graph capable of mapping non-obvious correlations between disparate assets (e.g., energy equity shifts driving crude oil moves).

Step 5: Proactive Push and On-Demand Pull

Deliver intelligence through two interaction modes: proactive push (displaying portfolio-relevant updates whenever the user opens the software) and interactive pull (LLM query resolution in 30 seconds for research that normally takes days).

When This Works (and When It Doesn't)

This pipeline works when you synthesize massive volumes of high-velocity public market signals with sensitive, siloed internal trading desk intelligence without violating client data privacy. As Sokhal points out, “I break the market up into all the different types of data which a trader looks at to decide what trade to put on and how to manage the risk.” Structured prices tell you what changed. Unstructured commentary and private vessel tracking explain why it changed. Unifying both into a private knowledge graph surfaces alpha that competitors miss.

It breaks down when applied to slow-moving markets or workflows without rich unstructured context. If your investment strategy relies on static quarterly filings, maintaining an n-dimensional graph creates unnecessary infrastructure overhead. The pipeline also fails if you skip local parsing. Routing confidential trading reports or internal tanker tracking through shared commercial LLM endpoints violates enterprise security protocols immediately.

What to Do With This

If you are building an intelligence platform for commercial logistics or freight brokers this week, apply this ingestion model directly. Do not route customer operational data into standard third-party APIs. First, separate public spot freight indices from your client's private broker rate sheets and email communications (Step 1 and Step 2). Deploy a local, open-weights vision-language model to parse messy carrier bills of lading (Step 3), tagging load origins and shipping lines against a strict domain ontology. Deposit these nodes into a local graph database (Step 4) so a rail delay in Chicago instantly connects to inventory risks in Atlanta. Finally, set up your interface (Step 5) to push relevant disruption alerts the moment a broker logs in, while letting them query contract exceptions in seconds.