Daloopa CEO Thomas Li: The Magic Isn’t in the Model
In the latest episode of Invest with AI, Brett and Khe sat down with Thomas Li, co-founder and CEO of Daloopa, who has been building AI into financial data collection since 2019. Brett's framing: he knows about three people who sat in a hedge fund seat and were talking about AI before November 2022, and Thomas is one of them.
For most of Daloopa's life, only about 5% of its database got touched by customers in a given quarter. They collect everything companies disclose, which means they never knew which 5% it would be. That number is now close to 100%, and data consumption has grown over 100x year on year. Thomas's read is that the unlock was not a new capability. It came from firms doing one simple thing: connecting MCPs instead of using chat as a chatbot.
The takeaway is not that MCPs solved the data problem. Thomas's argument is that an MCP is just a set of English instructions sitting in front of an API, which makes it only as good as the API underneath, the tables behind that, and whether the vendor connected them. Not all MCPs are created equal. The analyst who knows which pipes are worth turning on is the one who gets something out of an agentic workflow.
We get into:
Why only 5% of Daloopa's database used to get touched each quarter, why it is now almost all of it, and what a 100x jump in one year says about how research is actually changing
The unlock that was not a new capability, going from chat as a chatbot to connected MCPs, and why the users pulling thousands of times the average are almost always in a coding engagement
What an MCP really is, a long prompt in front of your API, and the four things that separate a good one from a brittle one: real data behind it, a searchable index, tables connected internally, and low latency
Why the biggest reliability gains are coming from data vendors rather than the labs, and how shipping skills alongside an MCP became the new version of the customer case study
The cannibalization debate, where putting everything in the MCP means customers stop opening your product, and why Thomas is willing to do it anyway
The water pipe framework for building a stack, pure enough for a science lab or just cold enough to cool machinery, and why you connect every pipe but only turn on the ones that solve your problem
Where multi-vendor stacks quietly break, because end-of-day pricing maps onto fundamentals cleanly but L1, L2, L3 and options do not, and no LLM is resolving a security master at runtime
Why the moat is the factory and not the tool, the Toyota six sigma comparison, and why incentive design at the 15, 30, 45 and 60 minute marks of earnings is part of what produces accuracy
The 13-week problem, finding the most painful week of a quarter and then the most painful minute of that week, and why the best workflow is an invisible one
What a knowledge graph is under the hood, folders and markdown files wired to each other, and why pre-doing the analysis improves output quality and drops token cost at the same time
Open weights versus frontier models, and why post-training lives or dies on the objective function, since math proofs are trainable and poetry is not
Whether AI can pass judgment, what factor models already quantified through PCA, and Thomas's argument that whatever is not quantifiable today is the new definition of judgment
Why Claude in Excel still cannot handle a real buy-side model, using operating leverage as the example, and why Daloopa is building its own AI-native add-in, still unlaunched, rather than wait for the labs
Eval-first product development, where every PRD starts with a set of evals and the questions often come from customers, and what changes by spring 2027 as the middle layer of the analyst job collapses
If you are building an agentic research stack at a fund, or trying to work out which data pipes are worth paying for, start here.
Chapters (Timestamps)
[00:00] Intro
[01:29] Founding Daloopa in 2019 to Take On the Data Oligopoly
[02:47] Accuracy, Latency, Trust: Why Human Data Collection Breaks
[05:45] From 5% of the Database Used to Almost All of It
[09:22] The Real Unlock: Chatbot to Connected MCP
[10:12] Not All MCPs Are Created Equal
[14:25] Where MCP Reliability Is Actually Improving
[16:16] The Cannibalization Debate
[17:54] Building Your Stack: Which Pipe Do You Turn On?
[21:02] One Pipe or Nine? Where Multi-Vendor Stacks Hallucinate
[24:31] The Moat Is the Factory, Not the Tool
[26:59] Going Deeper, and the 13-Week Problem
[31:30] How Funds Are Building Their Orchestration Layer
[33:29] Knowledge Graphs: Folders, Markdown, and Pre-Done Analysis
[35:44] Open Weights vs. Frontier: Where Post-Training Wins
[40:08] Post-Training Is an Objective Function Problem
[42:49] If Knowledge Work Isn't Verifiable, Is There a Ceiling?
[45:28] Can AI Pass Judgment? The Factor Investing Case
[51:40] The Race to the Middle: Quants and Fundamentals Converge
[53:18] Why Claude in Excel Still Can't Handle Operating Leverage
[58:47] Eval-First Product Development
[1:01:17] Spring 2027: Hiring, Training, and the Collapsing Middle
Want to actually build these workflows yourself?
The AI Accelerator is Fundamental Edge's 6-month cohort for investors who want repeatable AI workflows. Learn More below:
https://www.fundamentedge.com/ai-accelerator
Follow Invest with AI on:
Spotify: https://open.spotify.com/show/033xcEEovVViS7hIYwNuGZ
Apple Podcasts: https://podcasts.apple.com/us/podcast/invest-with-ai/id1896918892