Building AI Analysts for Wall Street: A Q&A with Compound Co-Founder Yijia Liang
Ahead of AI Engineer New York, The Daily Context spoke with Yijia Liang, co-founder of Compound, about building autonomous financial analysts, why general-purpose AI falls short in finance, and what it takes to earn the trust of high-stakes users.
The Daily Context: Tell us about yourself and Compound.
Yijia Liang: I'm EJ, one of the co-founders at Compound. We're building AI for Wall Street and finance, specifically private equity. We're building long-running autonomous analysts that behave and perform tasks the same way they would if you delegated the work to a human.
Previously, I was a vice president on a private markets team, investing in late-stage companies like Rippling, Databricks, and ClickHouse. Before that, I was an investor at Y Combinator's growth team, and I began my career at Bain in its private equity group.
I've spent a lot of time in buy-side finance, and I've found that the most interesting parts of the job aren't where people spend most of their time.
What's exciting about dealmaking is understanding a company, determining whether it's a good investment, and deciding whether to invest. But people spend most of their time building Excel models, PowerPoints, and Word memos, along with all the research and orchestration that goes into a deal.
That's what Compound is intended to help with. We're building much higher-quality outputs than you'd normally get from Claude or ChatGPT off the shelf, specifically for financial services.
The Daily Context: What will you be talking about at AI Engineer New York?
Yijia: The talk is about long-running autonomous analysts for finance and how we've approached that problem.
Recently, we've seen personal agents reach a kind of takeoff velocity because they've crossed a certain threshold of quality and productivity. But in finance, a lot of people in private equity and banking are still using AI as a chatbot. They're using ChatGPT or Claude to answer questions, or as plugins in Excel and PowerPoint.
That's very different from saying, "I'm going to delegate an entire workstream to you and come back 12 hours later."
That's what we're focused on building. The talk will walk through the technical details of how we're approaching that problem.
The Daily Context: When you set out to build an AI system that actually does financial analysis, rather than another general-purpose agent, what fundamentally broke?
Yijia: From day one, our focus was on how we could credibly produce ROI for the end user. In our eyes, that was the analyst or associate doing the grunt work. ROI means getting hours back in your day. Instead of you doing the Excel model or PowerPoint, Compound does it for you in exactly the same way you would have done it.
Initially, the models weren't good enough. We spent a lot of time figuring out how to get o3 or Sonnet 3.7 to work better with Excel. The models improved, but we realized three other things were broken.
First, the tools themselves. Excel, PowerPoint, and Word are built for humans, not agents, and certainly not cloud-based agents running hundreds of parallel processes.
Second, collaboration. So much of finance happens in small teams, with multiple people providing opinions and feedback. But AI has been so single-player. Even if you're giving an individual a few hours back, you're not solving the bigger problem of how to give hours back to the team or firm.
Third, context. Even with skills, memory, and dreaming, off-the-shelf models aren't perfect for finance because of the sheer number of unstructured files involved. Code is largely characters and Markdown files. Finance involves PDFs, Excel spreadsheets, and PowerPoints, which are much richer in information.
The Daily Context: What does good actually look like? Is it getting to the right number, showing the steps, or something more?
Yijia: There are two ways I think about good.
The first is reaching this almost magical threshold where someone looks at your output and says, "Yeah, that's what I would've done," and gives it a checkmark.
We're already seeing that in coding. Basically every software engineer I know is using AI to drive their coding. They look at the result, decide it looks good, and check it off. If we can produce financial models and PowerPoints that reach that same bar, that's great.
The second is whether AI passes the sniff test for judgment and critical thinking. That's where I don't think we're there yet.
Evaluating deals requires creativity, judgment, and self-iteration. When you work with an incredible human analyst, they often have next steps and follow-ups ready, even without being asked. They'll proactively address issues before you raise them.
AI isn't there yet. It produces something, and the implicit understanding is that there isn't necessarily a follow-up or next step.
When you work with a really good human analyst, they've usually already anticipated the three comments you're going to have. That's the level of quality we aspire to.
The Daily Context: Fast-forward five years. Will AI remain embedded in Excel, or will the entire interface for financial analysis change?
Yijia: Excel has survived for 40 years for two reasons. It's an incredible piece of software, and it functions as both a calculator and a canvas for financial firms to express ideas.
Some firms have 40 years of financial history stored in Excel. Unless you figure out how to map that information into a structured database, it's going to be difficult to move away from it.
Five years from now, I think Excel will still exist. But most firms will have moved toward having a data warehouse and will be migrating firm-level information out of Excel.
I think we'll see people relying more on HTML or another type of artifact that isn't Excel.
But people will still want auditability and traceability. How do you step into calculations and fully trust that a number is correct?
That's what Excel is great at. You can see the math, and the engine is deterministic. Whatever replaces it needs to offer that same trust.
The Daily Context: Do you think vertical AI is the future, or will general-purpose models eventually serve specialized industries like finance?
Yijia: I think about AI in terms of how it augments what differentiates one firm from another. For some firms, that's their human capital, the history of deals they've done, their relationships, or their platform.
You can accomplish that by building your own AI, using general-purpose tools, or using domain-specific solutions.
I strongly believe there will be vertical-specific solutions in finance, and multiple winners, because the category is so massive.
The reason is the unique nature of the data. Look at the most successful financial businesses before AI: Bloomberg, Capital IQ, and FactSet. They're built around supplying data to customers.
But what's becoming more interesting isn't just the data you provide, which will become increasingly commoditized. It's how you help firms harness their own data.
I think every firm will eventually have a data warehouse, a data lakehouse, and an ontological layer. I call it a "company brain."
The big winner will be someone who can operate that brain and become a real partner as firms deploy AI across their businesses.
The Daily Context: Financial services companies handle enormous amounts of sensitive information. How should they think about sharing data with AI models?
Yijia: Zero data retention needs to be table stakes. Model companies should not train on your data. That's our stance.
We have zero data retention agreements with every one of our model providers. We've chosen not to roll out new models if they don't have that protection available.
The next question is how you build valuable systems on top of that data. Firms will want to own those systems themselves.
You have to start with the most stringent version of security. Start from least privilege rather than most privilege.
The Daily Context: What's your advice for AI engineers building tools for high-stakes users?
Yijia: Two things.
First, you get one or two shots on goal before finance users form an opinion of your product. And that opinion is basically set in stone for a long time.
When Claude came out with its plugins and Opus 4.5, the narrative on Wall Street became that Anthropic had won the enterprise. Even when newer models came out from competitors, that narrative didn't change much.
Finance users treat AI products almost the same way they treat human analysts. They have incredibly high expectations. If you can't deliver the first or second time, they're going to write you off.
Second, you have to constantly reinvent. There's creative destruction happening in product development because AI is moving so fast.
Every week or month, we reevaluate what we've built. Is there an opportunity to rewrite it, make it better, make it faster, or even start from scratch because of how much AI has improved?
The Daily Context: What's something you wish more people would ask you about?
Yijia: I don't think people talk enough about eval construction in finance.
We're focused on the outputs that move the needle for customers. Our evals focus on entire workflows, from information ingestion and research gathering to writing and grading the outputs.
A lot of benchmarks evaluate whether the answer is correct, rather than evaluating the entire body of work.
But in finance, the body of work is actually the most important thing people evaluate.
We focus on high-quality outputs because that's what drives ROI for customers. How realistic our evals are compared with others in the market is a real differentiator for us.
The Daily Context: Is there anything else you want to highlight?
Yijia: We talked earlier about multiple people working with AI at the same time. I think that's going to be a big part of the future, especially for deal teams.
With our new product version, we've built that in natively. It's almost like a shared Slack channel where multiple people can talk with AI together.
Our belief is that you have to unify where people collaborate with the outputs they're working on.
That might mean multiple people talking in a channel about a financial model or leaving comments on a PowerPoint for the agent to incorporate.
It's a much more natural way of working. From an engineering perspective, it's also difficult to build a system that works well when taking feedback from five different people at once.
I think multiplayer and collaborative AI will be a major next frontier.