Brightwave - your autonomous investment research system 📈

CV Deep Dive

Today, we’re talking with Mike Conover, co-founder and CEO of Brightwave.

Brightwave offers an autonomous investment research system, helping active asset managers in public and private markets generate and act on investment ideas with greater speed and comprehensiveness. Leveraging a structured approach to agentic AI, Brightwave processes and synthesizes vast amounts of data—from SEC filings, earnings call transcripts and real-time news to high-signal, long-tail content from the public internet—allowing users to explore complex ideas quickly and with far less effort than a traditional investment research workflow. The platform differentiates itself by synthesizing a complete, interactive research report. Far from a question-answer style tool, Brightwave produces living documents that provide details on demand and allow users to dive deep into any topic. 

Mike has over 15 years of experience in AI and machine learning, with a PhD in complex systems. His career spans work at LinkedIn, SkipFlag (acquired by Workday), and Databricks, where he created Dolly, the first LLM to demonstrate that truly open-source models can exhibit the kinds of instruction-following behavior we’ve come to expect from ChatGPT. 

In June 2024, Brightwave announced their $6M seed round led by Decibel Partners, with participation from Point72 Ventures, Moonfire Ventures, and angel investors, including leaders from OpenAI, Databricks, Uber, and LinkedIn.

In this conversation, Mike shares how Brightwave combines document retrieval with synthesis to generate deeper insights, discusses the toughest technical challenges in building the platform, and explains why ownership is central to Brightwave’s team culture.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Mike 💬

Thanks for joining us Mike — Let’s start with some background on yourself and what led you to found Brightwave. 

Sure! I’ve been building AI and machine learning technologies for over 15 years, and I’ve worked on a wide range of machine learning teams, including tours of duty at LinkedIn, SkipFlag (a knowledge graph construction startup acquired by Workday), and most recently at Databricks. While at Databricks, I developed a language model called Dolly, which was the first to demonstrate that truly open-source technologies could exhibit the kinds of behaviors we see in LLMs like ChatGPT. Although it seems obvious now, back in April 2023 it was not clear this should be true, and our work played a major role in advancing the discussion around the democratization of AI technologies. 

My co-founder has an equally deep background—he was the CTO of LedgerX, a federally-regulated derivatives exchange and clearinghouse, and filed his first deep learning patent back in 2018. We met while he was leading a large-scale semantic search effort at Workday, and we share an abiding interest in machine learning and finance.

We see the markets as the ultimate manifestation of how humans allocate and negotiate value, which obviously is attractive from a commercial standpoint, but also means we’re solving an incredibly interesting problem. If you think about the job of an active asset manager, it’s to understand something about the world nobody else has seen to identify a mispriced asset. In the context of the totality of information relevant to an investment decision, it seems clear that language models can expand the limits of human cognition, and, fundamentally, transform how humans understand complex systems like the global economy or the financial markets. 

What’s the one-liner on Brightwave for the uninitiated?

Brightwave is building an AI-powered investment research assistant that helps financial analysts understand investment theses in greater depth and with greater speed. We have an agentic system that hardens and improves its understanding of a topic based on interactive input from an analyst, allowing them to go deep on ideas with a fraction of the effort it would have historically required.

Who exactly are your users? Is Brightwave primarily built for independent investors or for institutional investors? 

We’re seeing broad-based demand across the financial services sector, from buy-side teams at long-short hedge funds to mutual funds, private credit groups, private equity, and venture capital. Especially for users focused on public equities, our value proposition is very straightforward, as we operate over SEC filings, earnings call transcripts, real-time breaking news, and content from the public web to drive insights across synthesized research findings. As we continually grow our product feature set, we’re also finding that teams in consulting, corporate strategy, corporate development, and investor relations can all benefit from Brightwave.

Let's dive a bit deeper into how the product itself actually works. What data are you processing, and how do you handle it?

We treat this as a deep technology problem, not just a wraparound for commercial LLM providers. We’re able to leverage the Brightwave team’s extensive experience in AI and machine learning to train our own models, alongside serving and orchestration layers that specifically address the needs of a pretty sophisticated user base. We’re quite practical about the limitations of language models, and our approach to agentic behavior is focused on building “lane following” before full self-driving. We don’t believe the current foundation models can fully self-supervise autonomous research, so we take a more constrained approach that allows the user to direct the attention of the system interactively.

Think of it like a Directed Acyclic Graph (DAG) or a sequence of behaviors where non-deterministic judgments are made at each step. We circumscribe the workflow’s activities, allowing the model to make decisions it’s equipped for but without leaving everything to the model. With Brightwave, the user directs the system’s attention. For example, in our reports, users can highlight any passage in a Brightwave research report and invoke a library of LLM-powered analyses, directing the system to drill into areas of interest.

In terms of Retrieval-Augmented Generation (RAG), we treat search as a first-class citizen. Our approach focuses on search infrastructure that leverages semantically-aware representations of documents. For instance, we recognize that the executive preamble in an earnings call differs from the dialogue between an analyst and an executive, so we go beyond segmenting paragraphs to truly understand the meaning of documents. This approach leads to a higher quality of analysis and thought, which our customers have consistently told us is far superior to anything else on the market.

Users can also ingest their own content into the platform, and we will reason over that content with the same level of quality and meticulousness. Additionally, we pull material from the public Internet, as well as proprietary sources we've purchased access to. So, it’s a combination of their own data, public data, and our proprietary datasets.

Could you give us a better understanding of the competitive landscape you’re working in? And then, what sets Brightwave apart from an end user perspective?

There are a handful of companies in this space, and our users tell us we’re way out front on the basis of depth of thought and product experience. When I think of the value proposition for buyers, I think of it like Maslow’s hierarchy of needs. At the base, you have simple fact-finding and question-answering, essentially needle-in-a-haystack retrieval. For example, you can put a document into a large context window and ask a direct question, and the model will retrieve the answer—provided it doesn't require complex thought or calculation. 

The next layer up would be summarization, where given a document, Brightwave helps the user understand the gist of what it contains. We do all of that, but we see these as features vectoring towards commoditization, not standalone products. The real value comes in synthesis—connecting fact patterns across multiple heterogeneous documents. It’s about taking all the components like question-answering, summarization, and information retrieval and then tying them together to help users understand the complete story across many documents. Our approach introduces a lot of very interesting R&D challenges—the UX problem we’re solving is how to reveal the thought process of a system that has considered thousands of pages of material and present that to a human in a meaningful way—it’s highly non-trivial.

So, if synthesis makes connections that might otherwise go unnoticed, the next level of the pyramid is idea generation—identifying the most salient fact patterns and suggesting what kinds of opportunities or nuances they might entail. Finally, at the top of the pyramid is fully autonomous research, where the model not only retrieves and synthesizes information but observes its own work process and makes decisions about which threads to pull. Brightwave operates at the higher levels of this stack because basic fact-finding and retrieval are just table stakes—they’re not that complicated. The synthesis and self-directed organization of the system's own work product are what set Brightwave apart.

What are your customers’ use cases that you find particularly illustrative of Brightwave’s effectiveness? 

These are the top three areas where Brightwave shines: first, it helps users make investment decisions faster and with more conviction by making sure they aren’t overlooking anything crucial. Second, it enables teams to reason about massive volumes of content that would otherwise be impossible for a human to meticulously and effectively process. Third, it helps users solve the “blank page” problem, enabling users to generate ideas faster and evaluate factors they might not have otherwise considered. 

A concrete example from our own work: I started with a broad question about how Nestlé is positioned to capitalize on the secular trend towards more plant-based diets. Brightwave retrieved and analyzed thousands of pages of relevant material and highlighted that Southeast Asia has strong cultural inclinations toward plant-based meat alternatives. In terms of idea generation, Brightwave identified that Singapore is the only municipality with regulations permitting the retail sale of lab-grown meat. Pulling on that thread further, Brightwave found that JUST (the company behind plant-based mayo and egg substitutes) had recently secured something like a quarter million liters of bioreactor capacity in Singapore to produce lab-grown meat and that Nestlé may face production headwinds in this very niche sub-sector of a sub-sector.  

From a high-level, macro-thematic starting point, the system, with very limited supervision on my part, drilled down to super-granular insights on bioreactor capacity in Singapore. I like how this illustrates how Brightwave supports both breadth (processing large volumes of content) and depth (exploring specific avenues based on the user’s inquiries), and is representative of how our users put the system to work.

You mentioned that this was a deeply technical problem. What was the hardest technical challenge in building Brightwave?

The hardest challenge has been dealing with the cumulative effects of methodological complexity in a system like ours. Imagine a physical system with a single joint that has just a little bit of "play" in it. On its own, you might not notice much, but when you have eight joints, all those tiny imperfections add up to create significant travel in the system. Brightwave involves multiple machine learning and generative subsystems—think of a Directed Acyclic Graph (DAG) with many different LLM calls, retrieval events, and self-critiques happening simultaneously. The challenge is optimizing all these elements jointly: how models critique their own work, how they process documents, what retrieval models are doing, and how fine-tunes affect the system. All these components must be measured and co-optimized. We can’t perform gradient descent on the entire system, and so we’ve had to develop new methods for measuring and improving system performance holistically. This is where the teams’ experience on consumer-web machine learning teams comes into play. In those environments, machine learning subsystems often consume the outputs of other models. While each model might be differentiable on its own, the entire system isn't, making optimization at scale incredibly difficult.

The core of the challenge is measurement and evaluation. It's about designing a wraparound "compass" that consistently points to true north, even though each system has its own nuances. It involves synthesizing evaluation data, human annotators, using LLM supervision to determine if a set of results is satisfactory, and, of course, classical statistical methods for measuring model performance for individual components of the system. Essentially, it's solving an integration testing problem, where every change must be evaluated not only at the atomic level but also at the composite level. You push a battery of examples through the system to ensure that, in the aggregate, it still behaves as expected. This is why so many other companies are breaking on these rocks—it’s a deeply complex, multi-layered engineering and research challenge that requires careful, continuous measurement at every level.

Are there any particular trends in AI that excite you, and how do those trends influence the way Brightwave will evolve over the next year?

Yes, there are several trends in AI that excite me. One area I'm particularly interested in is reinforcement learning-based planning algorithms that treat LLMs as just one step in a sequence of actions. Right now, agentic workflows tend to rely on something like greedy search, and I think the next wave of capabilities will come from models that jointly co-optimize planning and generative inference alongside the orchestration systems and serving unit economics that allow for speculative execution of many sub-plans. 

It's also unclear to me whether we're approaching saturation on model quality yet. I do believe we'll see several more rounds of improvement, but it's fascinating that so many independent labs have reached similar benchmark scores without anyone having an extraordinary breakthrough, despite the massive economic incentives for doing so. That remains the billion-dollar question.

At Brightwave, we're skating to where the puck is going. We're building a system that's at the razor's edge of what the current state of the art is currently capable of doing, and Brightwave will improve as foundation models improve. However, we're not betting the farm on a massive breakthrough in model quality. Instead, we're focused on solving real problems today, with the expectation that tools will continue to improve incrementally.

Tell us a little bit about the team at Brightwave. What kind of culture are you building, and what kinds of roles are you hiring for?

We are definitely hiring, and our talent philosophy can be summed up as "operators only." We believe small teams of exceptional, opinionated people who know what excellent work looks like will outpace much larger groups. Ownership, humility, and craft are what we prize, with ownership mindset defining how we show up every day. We hire people who intrinsically understand, because of their upbringing, life experiences, whatever, that they have an inalienable right to choose how they respond to their environment. It’s about having an internal locus of control, and recognizing that your ability to take an active role in navigating uncharted waters is the key determinant of the outcomes you experience in life. That’s the hardest thing to hire for, but we’re lucky to work with a group of people who embody these principles to a tee. They have extraordinary professional experiences, whether it’s publishing in NeurIPS or tours of duty at places like Meta, Goldman Sachs and UBS, but the thing that’s special about Brightwave is the culture. We've hired exceptional people—interesting, empathetic, intense. People want to work here. We've got folks turning down offers from the top companies in AI to join us. We've built something special.

As for roles, we fit the job to the person, not the other way around. If you’re an operator, we’ll find a place for you to do the best work of your life. We’re hiring search engineers, AI/ML engineers, systems engineers and designers to build an intelligent system like nothing else that’s ever existed before.

Conclusion

To stay up to date on the latest with Brightwave, learn more about them here.

Read our past few Deep Dives below:

• Zilliz Cloud is powering vector search for the enterprise 🔋

• Kolena aims to redefine quality standards in ML 🎛

• Pinecone Serverless is turbo-charging your time-to-prod ⚡️

• Cleanlab is automating data curation at scale 🌐

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email ([email protected]) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs 

Event Spotlight

Also, if you’re in San Francisco this weekend, join us for SHACK15 Conversations / 2040: A Silicon Valley Satire Author Pedro Domingos! 

Join us as we welcome Pedro Domingos – a renowned AI researcher and tech industry insider – in conversation with SHACK15’s founder Jorn Lyseggen about the development of AI. The jumping off point will be Domingos’s fiction debut – 2040: A Silicon Valley Satire – a book that imagines a presidential election unlike any in U.S .history. The year is 2040 and an AI named PresiBot is facing off against a fake Native American chief seeking to abolish the United States. PresiBot is hallucination prone and buggy, KumbAI's brash CEO Ethan Burnswagger has lost the override button, and the fate of American’ democracy hangs in the balance, but there are bigger fish to fry because the startup is almost out of funding.

Piercing the zeitgeist around AI and set in a dystopian San Francisco in a near future we can all too easily anticipate, it features characters, entities, and incidents whose resemblance to actual ones may or may not be purely coincidental. The book raises important questions about what AI really is, how the tech industry works, where our deepening polarization might lead us, and―most important―how to break out of this cycle. Copies of 2040 are available at Book Passage in the Ferry Building.

Register here.