← back
Trust, but Verify: Knowledge Agents for Finance Workflows - Mike Conover
Takeaway
Financial knowledge agents need verticalized scaffolding, multi-pass self-verification, and human-in-the-loop nudges—chat alone won't do trust-but-verify work.
Summary
- Brightwave (Mike Conover, ex-Databricks Dolly creator) builds research agents for due diligence over thousands of pages of filings, transcripts and vendor contracts.
- Non-reasoning LLMs do greedy local search; winning systems will end-to-end RL over tool-use traces so locally suboptimal calls yield globally optimal outputs.
- Design patterns: decompose the human analyst's process, distill findings, then enrich/error-correct in a separate call (models are too credulous in-context).
- Argues vertical product scaffolding plus human oversight is durable because non-experts won't invest 1000+ hours becoming prompting wizards.
agentsfinancerag
Original description
Join us for a deep dive into the engineering and interaction design patterns that power the automated creation of high-signal, information-dense investment research reports. Distilling accurate, actionable insights from vast, multimodal data sources requires a thoughtful approach to both the underlying models and the product affordances that give users granular visibility into a system’s reasoning across thousands of pages of material. We’ll explore how these design considerations foster trust, clarity, and effectiveness in modern financial workflows. 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 Mike Mike Conover is the CEO and co-founder of Brightwave, the AI research and diligence platform purpose-build for investment professionals. A pioneer in the field of artificial intelligence, prior to Brightwave Mike led LLM engineering efforts at Databricks where he created Dolly, one of the first open-source models to exhibit instruction following behavior. Mike has built AI systems & products for more than fifteen years, including tours of duty at Workday as Director of Financials Machine Learning, SkipFlag (acquired by Workday), and LinkedIn. He has a Ph.D. in complexity science from Indiana University, and his work has been featured in Bloomberg, The Wall Street Journal, The New York Times, TechCrunch, Hacker News, Nature Communications, MIT Technology Review and on NPR.