Brightwave

Brightwave builds AI systems for investment research and professional analysis. Co-founded by Mike Conover and Brandon Kotara, it developed a financial-research assistant that combines retrieved sources, structured relationships and specialized reasoning subsystems. Its subsequent product generations added autonomous research workflows and dedicated agent computing environments capable of producing reports, presentations and financial models.[2][1][6][7]

The founders brought experience in language models, information retrieval and financial infrastructure. Before Brightwave, Conover established and led open-source language-model engineering at Databricks, where he worked on Dolly. Kotara had served as CTO of LedgerX, a federally regulated derivatives exchange. In June 2024, Conover was Brightwave’s CEO and identified Kotara as its CTO; he also described Kotara’s earlier work leading semantic search over hundreds of millions of résumés and job listings at Workday.[2][1]

Conover described the assistant as a “partner in thought”: a way for finance professionals to investigate themes and their implications. In the June 2024 interview, he said hedge-fund equities teams used it to deepen investment theses, wealth managers used it to prepare for client conversations, and investor-relations and corporate-strategy teams used it to understand markets. As a concrete example, he recalled Brightwave tracing investment opportunities around AI demand into energy and then long-duration energy storage. Humans remained responsible for assessing whether those theses fit their strategy and knowledge of management teams.[1]

Brightwave’s architecture reflected Conover’s reduced confidence that long context windows alone could produce detailed synthesis. In June 2024, he explained that a whole-book summary could lose detail that chapter-by-chapter analysis preserved. Brightwave instead decomposed documents and reasoning tasks into specialized subsystems. Retrieval considered document structure and the time intent of a question: breaking news required a different selection of material from an account of the previous 18 months of Federal Reserve policy.[1]

Conover described retrieval-augmented generation as a way to ground reasoning in facts available at inference time, while fine-tuning specialized repeatable behaviors rather than supplying new factual knowledge. The system also used classical machine learning where suitable. Its evaluation methods included comparing different models’ conclusions, training models to assess whether primary sources supported a generated claim, and calibrating automated evaluations against human domain-expert annotations. Users could request additional checking of surprising findings.[1]

Structured extraction complemented document retrieval. In June 2024, Conover described investment in extracting relationships in a single pass so that documents would not need to be reanalyzed for every query. By October, Brightwave described a proprietary knowledge graph built from SEC filings, earnings-call transcripts, breaking news and public internet content, with a reasoning system layered above it. The graph represented factors including supply-chain relationships, M&A, governance changes, regulatory approvals, intellectual-property litigation and cybersecurity events.[1][3]

On June 11, 2024, Brightwave announced that it had secured $6 million in seed funding, led by Decibel Partners with participation from Point72 Ventures, Moonfire Ventures and angel investors associated with OpenAI, Databricks, Uber and LinkedIn. The company reported that, after four months of growth, its customers included funds and firms managing more than $120 billion in assets. That figure described customers’ aggregate assets under management.[2]

On October 29, 2024, Brightwave announced the close of a $15 million Series A, again led by Decibel, with OMERS Ventures participating. The round brought its reported total funding to $21 million. Brightwave also reported fourfold revenue growth in the four months since its seed announcement, without disclosing an absolute revenue figure. It said the financing would support engineering and commercial hiring, strategic data partnerships, and research and compute for its knowledge graph and reasoning engine.[3][4]

Brightwave expanded its data offering through a Quartr partnership announced February 11, 2025. The integration made Quartr’s public-company earnings transcripts, investor presentations, filings and other material available alongside customers’ proprietary data. It was announced as immediately available to all Brightwave customers, supporting both private-market diligence and public-market research.[5]

On August 13, 2025, Brightwave launched Research Agents, a multi-agent system controlled from a unified interface. Its announced workflow began with a user-defined objective: the system generated a plan, selected tools and executed the task, then submitted the output to an LLM judge for assessment against the user’s requirements and revision when needed. Tasks included full-length reports, evidence-linked tables and visualizations, document refinement and primary-source exploration. Users could inspect interim results and adjust the agents’ focus while work continued in the background.[6]

The March 10, 2026 release replaced Brightwave’s shared execution layer with dedicated agent sandboxes. Brightwave described each environment as having its own filesystem, controlled internet access and persistent state throughout the conversation, using gVisor isolation. Agents could install tools, execute code and retrieve external data. Research, document and Excel agents coordinated as peers through structured messages, exchanging findings and flagging inconsistencies.[7]

Spreadsheet generation marked a concrete change in product scope. In June 2024, Conover had called creating Excel spreadsheets an explicit non-goal, while emphasizing that Brightwave already performed quantitative reasoning. He considered subtle spreadsheet errors difficult for users to detect and financial modeling closely tied to an analyst’s personal understanding of a company. By March 2026, Brightwave was announcing Excel agents and multi-sheet financial models alongside Word reports and PowerPoint presentations.[1][7]

In a March 18, 2026 engineering account, founding machine-learning engineer Thet Naing explained how Brightwave connected generated claims to exact source locations. Ingestion attached character offsets for text, page bounding boxes for rich documents, cell ranges for spreadsheets and paths for JSON or XML. That metadata traveled through retrieval into agent context. Separate viewers resolved citations to highlighted passages, cells or structured-data nodes; text citations required offset transformations to account for changes introduced by rendering.[8]

Naing described two validation layers: rules checked that source positions were structurally valid, and LLMs checked whether the cited material supported the generated claim. Production monitoring tracked resolution failures, citation density and performance by document format. His account reported nearly a million citation evaluations over the preceding 30 days—an evaluation-volume figure rather than an accuracy rate.[8]

Software engineer Nick von Pentz described Brightwave’s prompt-injection containment design on April 15, 2026. Real integration credentials stayed outside agent sandboxes: agents used dummy keys that a separate proxy replaced on outgoing requests. Network-level iptables rules enforced proxy routing rather than relying on cooperative environment settings. The proxy restricted destinations and checked for the token issued to the sandbox, addressing attempts to reach an allowed service using an attacker’s credentials. These were infrastructure controls intended to bound a compromised agent’s capabilities, rather than depend on the model recognizing malicious instructions.[9]