Here’s an unordered list of short, concrete, and hard problems we’re looking to solve in the next year to transform financial markets.

1. Learning by Analogy:

Prompt engineering shouldn’t be a source of user error. We want AI systems to learn how we work rather than us learning how they work. It should be seamless for an associate to upload work they’re proud of — a deck, memo, or model — and ask an AI to replicate it on a different topic. "Emulate my deck on industry X for industry Y given the recent events around Z" should result in analogous work.

Bankers are efficient workers: they start from existing work and adjust. By building AI systems that learn by analogy, we can emulate those workflows. 

Compiled Devin steps from our friends at Cognition Labs

What makes this problem super hairy is the variance in user input. Structured outputs are quite tractable. When writing emails, base frontier models do a good job, provided they're given an email as an example. When you're given multiple different types of documents and asked to recreate the bundle on a new subject, the problem suddenly requires intuition, domain knowledge, and interpretability that those models lack.

Promising Ideas:

• Quickly configurable/specialized embeddings and reranking.

• Building Compilers: Similar to how compilers have a 1:1 mapping between assembly/LLVM and a programming language, we can generate mappings from financial document and update relevant variables.

• Novel research into understanding when an LLM needs to go deeper / reverse course on a financial topic, and when not to.

• Novel research into financial LoRA adapters - especially combined with new research in long-ICL.

2. Exact Context:

At Rogo, we've built an expansive search engine over most financial documents, ingesting news, web articles, company filings, and structured market data every minute. Finding exactly what's relevant, quickly, is paramount at scale.

Atop this mountain of data we've built a knowledge graph that supports efficient retrieval, and we need to figure out how to scale it to quickly answer any question people have about finance.

Robustness at scale truly is the problem here — our users are just as interested in the unit economics of pool cleaning companies as they are in the peak sales of Phase 1 oncology targets. We believe that models will eventually support retrieval, but until then, we’re in the business of building tools that agents can invoke to give precise answers.

We are incredibly excited about the future of retrieval. There are so many paths the technology can and will take, beyond embeddings, keyword search, and graph lookups. Trading indexing cost in exchange for query-time speed and accuracy, what's the most efficient way to search?

To answer this, I’d have to:

1. Find Pfizer's Phase 1 oncology drugs and understand what they do.

2. Understand what is missing from their R&D pipeline compared to their peers.

3. Understand how much cash and debt Pfizer can raise to acquire those companies, and how much each potential acquisition has raised.

4. Synthesize all of the above information to provide a direct answer.

Unlike many traditional search engines, our users tend to ask complicated, vague, multi-hop questions. Simple RAG and search architectures only work for single question retrievals, so we have to think from first principles and design promising systems that will perform at scale.

Promising Ideas:

• Social Graphs: By integrating with Microsoft Teams, Outlook, and AzureAD (or their non-MSFT equivalents), we can learn a lot about what documents are relevant to a specific banker. How can we use the internal social network of a bank to improve search?

• Custom, Interpretable, and specialized reranking for financial document search.

• Training encoders optimized for KG hops: given a query and a set of entities/information parsed so far, where should we hop next?

• Agential reasoning with tool calling: having an LLM agent capable of collecting information across a set of appropriate tool calls come to a conclusion from the results of those calls. Avoid having it abruptly end or run in loops.

• Infinite context and near-perfect financial market analysis.

3. Instant PowerPoint (file) Edits:

Bankers want to focus on finishing their work as fast as possible. Ideally, they should be able to automatically generate a slide deck, then iterate on it by chatting with the file to make edits and pull in fresh, relevant financial information. Just by chatting with Rogo, it should be possible to do anything you do in Excel, PowerPoint, and Word, all while having instant access to the world's most powerful financial search engine.

We don't think PowerPoint plugins are the UX to do this. Instead, we're building a way for bankers to never have to open up PowerPoint unless they want to.

An early prototype

Creating an AI system with the same intuition as a banker for PowerPoint is a tall order. But, we have a good intermediate representation that we’re quite happy with, which helps us encode intent.

Early experiments with frontier models in Rogo has led us to the following roadblocks:

• Models are lazy and often will give up instead of writing significant changes.

• Only writing diffs leads to significantly worse performance.

To compensate, we currently have to rewrite the entire intermediate representation of a slide with each edit, resulting in potentially thousands of tokens with every iteration of a file. This comes with the benefit of thinking with more tokens, as the model has more forward passes to sample correctly. Forcing us to only focus on diffs means we have fewer time to think.

This solution demands a hyper-fast inference infrastructure— at thousands of tokens per second — along with some incredibly intuitive and clever user design.

How might you trade off a more expensive, intelligent edit and maintain very fast inference? We might need to run the model in a background session, with some clever stream handling and UX along the way.

Promising Ideas:

• Given 15 minutes of events from a Rogo PowerPoint editing session, how might you predict the next minute?

• Speculative decoding for rebuilding a file from diffs, akin to our peers in code generation.

• Using Anthropic’s Artifact paradigm to display progress to the user.

• Working with RLHF providers to guide models to intuit what bankers mean in their requests.

• Understanding the relational aspect of Excel and PowerPoint more deeply — there’s so much metadata in each that can be extracted to understand intent.

4. Interpretable AI through Sentence-Level Grounding

Citations are highly optimized and weighted tokens to be sampled based on how the model is trained to construct them. But they're not quite "citations," as we don't know exactly which tokens were used to generate a specific claim. The answer to which sentences affected the generation of a claim ahead of them lies within the attention layers.

Promising ideas:

• Using sparse auto-encoders to identify which features significantly impacted generation.

• Understanding the extent answers are sycophantic instead of truthful.

• Build an intuitive UX around understanding the extent one source impacted the generation compared to another. This allows us to highlight bias in our AI models to certain types of sources, responses, and prompts.

—————

If you read this far, and you're interested in the problems we are interested in, then please consider joining us! The Rogo team is small and talent-dense, and there are a number of reasons why we think you would love to join us:

• People love using Rogo. We are a rare vertical GenAI company with real usage, power users, and true ROI.

• AI for Finance is a huge market. Generative AI is replacing human work, and financial professionals get paid more than anyone on earth. The commercial opportunity is a no-brainer.

• You will work with very smart people. Everyone at Rogo is deeply passionate and incredibly smart.

• We work hard because it's fun. Every day brings a new challenge, a new piece of feedback, and a new idea. We work very hard and are having so much fun on the journey to bring our vision to life.

Send an email to careers@rogodata.com, apply here, or send a direct message to one of our founders.