← back
Scaling Agents for Gen AI Products - Anju Kambadur, Bloomberg Head of AI Engineering
Takeaway
Finance-grade agents require semi-autonomous architectures with mandatory guardrails and tool APIs as rigorously documented as numerical libraries.
Summary
- Bloomberg's 400-person, 50-team AI org pivoted from training BloombergGPT to building atop open-source/frontier LLMs after ChatGPT.
- Daily scale: 400B structured ticks, ~1B unstructured messages, millions of documents over 40 years of history.
- Earnings-call summarization product required mlops remediation workflows and circuit breakers because outputs are published, not chatted.
- Agentic architecture is intentionally semi-agentic: autonomous reasoning paired with hard-coded guardrails (e.g., no financial advice, factuality checks).
- Stresses BLAS-style well-documented tool APIs as the foundation for fast agent iteration; non-negotiables are precision, comprehensiveness, speed, and transparency.
bloombergenterprise-agentsguardrails
Original description
Recorded live at the Agent Engineering Session Day from the AI Engineer Summit 2025 in New York. Learn more at https://ai.engineer and purchase tickets to our next event, the AI Engineer World's Fair, in SF June 3 - 5 here: https://ti.to/software-3/ai-engineer-worlds-fair-2025 About Anju Dr. Prabhanjan (Anju) Kambadur is the Head of the AI Engineering group at Bloomberg, which consists of 400+ researchers responsible for building financial solutions, including search and conversational systems, high-precision extraction pipelines, and time-series forecasts, as well as the firm’s underlying AI infrastructure. The members of his group are active in the academic community, where they have published more than 100 peer-reviewed papers in the last three years. Anju is one of the authors of the BloombergGPT research paper. He was also recognized on Insider's "AI 100" list of the top 100 people in artificial intelligence in 2023. Before Bloomberg, Anju was a research staff member at IBM Research’s Thomas J. Watson Research Center, where he worked on problems in machine learning, such as matrix completion and sketching, sparse coding, genome-wide association studies, temporal causal modeling, and high performance computing.