AI Engineer
← All speakers

Related work and projects

  • ChatGPT for Financial Services ↗

    In his October 14, 2026 talk, Kim identifies himself as the lead of ChatGPT for Financial Services at OpenAI, initially focused on investment banking and equity research. He describes his team's development work with Morgan Stanley and Evercore as design partners: financial-data access, citation viewers exposing source passages and tables, interactive financial charts, and spreadsheets and presentations using firm-approved templates. He separately describes an ongoing finance-specific post-training effort targeting financial reasoning and the quality of Excel models and PowerPoint presentations.

    openai.com · Source ↗
  • PitchBook integration with ChatGPT for Financial Services ↗

    Shouyi Poon credits Kim among the OpenAI collaborators on the PitchBook partnership. The accompanying PitchBook announcement describes access to company, investor and fund data through the Essential connector inside ChatGPT for Financial Services, with citations back to the source. The evidence establishes Kim's participation in the partnership alongside other OpenAI and PitchBook contributors.

    linkedin.com
  • Gemini 3 is live in Rogo ↗

    Article authored by Kim on November 18, 2025, announcing Gemini 3 Pro availability in Rogo. He describes the team's early-access collaboration with Google's Gemini post-training group, an architecture that separates retrieval, verification and reasoning infrastructure from the underlying model, and custom-trained evaluation models assessing usefulness, source credibility, formatting, concision and factual correctness. The engineering practices and collaboration are attributed to the Rogo team.

    rogo.com
  • GPT-5 integration in Rogo ↗

    Kim's personal post establishes his collaboration with Sherwin Wu and others at OpenAI to launch GPT-5 within Rogo. The August 7, 2025 launch article, authored by Gabe Stengel, describes testing on financial workflows including valuation analysis, precedent transactions and pitch-deck generation. Rogo evaluated answer quality separately from source attribution, assessing whether claims requiring citations had accurate and complete support. Kim is credited with launch participation; the article's authorship belongs to Stengel and its evaluation methods are attributed to Rogo.

    rogo.com · Source ↗
  • Rogo applied AI and financial-agent development ↗

    Kim announced joining Rogo to lead applied AI, identifying reinforcement learning and language models as the focus of his financial work. Rogo's company post subsequently identifies him as vice president of engineering and credits his leadership in improving its agents and expanding its product offerings. These sources establish contributions through his Rogo employment and leadership roles; they do not establish personal ownership of the company or its products.

    linkedin.com · Source ↗
  • BERT deployment at scale in Google Search ↗

    Gabriel Stengel's Rogo hiring announcement credits Kim with deployment work bringing BERT to Google Search at scale. It connects that work with his earlier MIT metamaterials and NASA robotics background, and identifies experience in multi-step reinforcement learning, specialized retrieval and large-scale computation. The supported contribution is deployment work at Google; the source does not establish authorship of BERT or its research paper.

    linkedin.com
  • MIT research on lightweight, reconfigurable composite structures ↗

    MIT News's August 15, 2013 report names Kim as an undergraduate participant in a project developing lightweight structures assembled from small interlocking composite components. The structures could be disassembled for repairs or reassembled in different configurations. The report credits Kenneth Cheung and Neil Gershenfeld as authors of the associated Science paper and names Kim among the project participants, without attributing paper authorship or independent invention to him.

    news.mit.edu

Bio, Work & Ideas

Joseph Kim

Conference affiliation: OpenAI

Joseph Kim leads ChatGPT for Financial Services at OpenAI, building AI tools for investment banking and equity research. His work combines language-model engineering with the practical demands of financial analysis: usable outputs and evidence an analyst can check.

As an MIT undergraduate in 2013, Kim contributed to research on lightweight composite structures. The approach used small interlocking parts to make structures that could be taken apart for repairs or reassembled in a different configuration. He later worked on robotics at NASA and helped deploy the BERT language model at scale in Google Search.

Kim joined Rogo to lead applied AI, focusing on reinforcement learning and language models for financial work, and later served as its vice president of engineering. He collaborated with Sherwin Wu and others at OpenAI to bring GPT-5 into Rogo. The August 2025 launch followed testing on workflows such as valuation analysis and pitch-deck generation. Rogo evaluated answer quality separately from source attribution: it compared responses for usefulness and presentation, and assessed whether claims needing citations had accurate and complete support.

In his November 2025 article on Rogo’s Gemini 3 Pro release, Kim described the team’s collaboration with Google’s Gemini post-training group to shape model behavior for financial use cases. He also explained Rogo’s model-agnostic architecture: retrieval and verification were kept separate from the choice of underlying model, reducing the engineering work needed to test and deploy replacements. Custom-trained evaluation models assessed whether the resulting answers were useful, factually accurate and supported by credible sources, while also checking formatting and concision.

At OpenAI, his team developed ChatGPT for Financial Services with Morgan Stanley and Evercore as design partners. The product brings financial data into ChatGPT. Analysts can inspect the source passages and tables behind an answer, explore financial charts, and create spreadsheets and presentations using their firm’s approved templates. Kim also worked on the PitchBook partnership, which brings private-market intelligence into the product. Alongside this product work, his team is pursuing further model training for financial reasoning and the quality of Excel models and PowerPoint presentations.

  • Reduce the cost of review. Kim’s threshold of usefulness ties AI adoption to the effort required to check and repair its output. If reviewing a financial analysis takes as long as producing it yourself, improvements in the model offer little practical benefit. Once review takes substantially less time, regular use and delegation become worthwhile. He argues that this threshold can make gradual gains in capability feel like a sudden change in the work people are willing to delegate.
  • Make delegation familiar. Kim sees persistent agents as a way to reduce the coordination burden of using AI. Such an agent would remember a goal and relevant context, keep working between prompts, and return with progress. Users could assign a project and ask for updates in a familiar way, making the system more accessible to professionals who have not learned elaborate prompting workflows.
New York event photos

Read the topics behind these talks

1 conference talk

Loading more talks and highlights…

References