The Recommendation Engine for Commodity Traders — Anekha Sokhal, Moshi
Summary
Moshi, the product discussed, is an AI-powered recommendation engine designed for commodity and energy traders. It aims to solve the problem of information overload by processing massive volumes of diverse data—including structured time series, unstructured news, alternative data (like X/Twitter firehose), and proprietary user data. The system integrates these inputs into an N-dimensional graph, providing real-time, actionable insights and recommendations that connect seemingly disparate information points, moving beyond traditional dashboards into predictive workflows.
Key takeaways
-
The Scope of Data Processing
Moshi processes data volumes that are immense, noting that the system can learn hundreds of new things in the time it takes to say a single sentence. Data types include structured time series (futures markets), fundamentals, and unstructured/alternative data (news, audio, X data).
-
The Connected Graph Architecture
0:10
The core innovation is building an N-dimensional graph that connects different data points. This allows the system to identify second and third-order derivatives—for example, linking an energy equities company's activity to an indirect impact on oil prices—which a standard news stream would miss.
-
Addressing Data Conflict and Trust
0:16
The system is designed to handle conflicting sources (e.g., Reuters vs. WSJ reports on shipping activity). The disagreement itself is treated as a point of data, and the platform provides a confidence score to help users assess the reliability of information.
-
The Recommendation System Paradigm
0:10
Moshi is characterized as a blend of Bloomberg's functionality and Netflix's recommendation system. The goal is to push relevant data to the user, making the platform indispensable and naturally integrating into the user's workflow.
Technical details
-
Data Types and Ingestion
0s
The system ingests three primary data categories: 1) Structured data (time series, technical indicators, moving averages); 2) Unstructured data (news, audio, X data); and 3) Private/Proprietary data (user-specific subscriptions or internal company data, such as LNG ship counts).
-
System Architecture and Modeling
0s
Data cleaning, transformation, and statistical analysis are performed before the data is mapped onto an N-dimensional graph. The output is a software interface that pushes real-time insights related to a user's portfolio, enabling LLM-powered answers in seconds.
-
Handling Messy Data
0s
The platform utilizes a vision language model (VLM) to process and structure data from messy sources, such as uploaded PDFs, breaking down the content and building it into the core database.
-
Model Ownership and Integration
0s
A key architectural consideration is data sovereignty. The speaker emphasizes the importance of owning the data processing models (e.g., building proprietary VLMs) rather than relying solely on third-party vendors, especially when dealing with sensitive client data.
Mentioned resources
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.