Rogo CTO Tumas Rackaitis says Wall Street runs on relationships between companies, so an AI analyst must map those connections through a knowledge graph that tracks funding rounds and acquisitions.
Tumas Rackaitis argues PowerPoint generation is a different modality from text, images and video, because banking decks must match each partner's or managing director's preferences, and Rogo learns them from precedent decks.
Rogo contrarily searches over an LLM's internal attention layers instead of standard embeddings, which Tumas Rackaitis says compress too much information and are technically reversible, at a high GPU memory cost on NVIDIA hardware.
Rogo builds Wall Street's AI analyst on relationship graphs
Rogo is building an AI analyst for finance by treating Wall Street as a web of relationships between companies. It pairs a knowledge graph with deck generation that learns a bank's style from precedents, and with a single place for private and public data. Its contrarian bet is searching a model's attention layers rather than embeddings.
Takeaways
An AI analyst needs a map of company relationships. Rogo built a security master that maps spoken names, such as Instacart, to the right company, then extended it into a graph of funding rounds and acquisitions. Tumas says Wall Street is a web of company relationships, and an AI analyst must understand them.
Slide generation is a different task from text. Banking decks follow the narrow preferences of a partner or managing director, and a miss means late nights of fixes. Rogo used finance experts, including people from Blackstone, to model those preferences. Tumas calls a deck the most condensed way to show value.
Learning from precedent decks beats rule-based automation. Bankers normally build decks from precedents, and a good AI system learns from examples quickly. Earlier tools built on UiPath and rule-based systems never really worked, while the new approach copes when small things go wrong. Logos stored in the graph and checked with computer vision save hours.
One general platform beats many embedded chatbots. Bankers do not need a separate chatbot in each application, such as an Excel plug-in or an expert interview platform. Rogo wants one place connecting a firm's private data to public information on private and public companies. Chat responses become units of work for automation.
Searching attention layers beats embedding-based retrieval. Standard retrieval runs an embedding model over files and returns documents similar to a question. Tumas says embeddings compress too much and are reversible given a convex optimization problem. Rogo searches an LLM's internal attention layers instead, costing GPU memory on NVIDIA hardware but working well.
Rogo's four-layer stack
Data layer. It crawls and links information to form a financial search engine.
Search layer. It indexes the data and builds the knowledge graph and the AI-native security master.
Chat layer. It answers questions in a ChatGPT-style request and response pattern.
Agentic layer. It generates PowerPoints and automates workflows from start to finish.
Tap a name to go there · tap a line to see why they are linked