I'm running an all-AI team of seven programmers and a manager. Why does it work? Because an agentic coding team is both a software architecture problem and a team management problem — and the same patterns that work for human engineering organizations work here.
At Nomis, we routinely build sophisticated neural networks we know will never hit production — specifically so we can have the right conversation with the C-suite about why the logistic regression is the right answer.
Rapid prototyping of a voice-enabled mortgage application system and lessons learned about compound AI systems.
Moving beyond binary recommendations to AI systems that provide arguments, alternatives, and contextual insights.
Why effective AI usage requires management skills — setting expectations, providing context, and giving constructive feedback.
Building AI systems that proactively surface insights for analysts rather than replacing human expertise.
Focusing on practical, value-creating AI applications — data analysis, recommendation engines, and decision support — rather than revolutionary claims.
Closing the gap between AI belief and execution — 85% of banks see transformation potential, only 10% have implementation roadmaps.
Why banks experienced 50-60% deposit betas when expecting 20-30%, and what to do about it in future rate environments.
Strategic considerations for managing deposit costs in a falling rate environment after years of underestimated deposit betas.
How sophisticated technology, including machine learning, can automate model management processes facing 3-5x volume increases.
A proposal for Validation to take on dual roles as engineer and quality control to streamline the model development process.
Using industry CD-Share as a rate proxy to add decades of rate sensitivity to deposit models trained on ultra-low interest rate periods.
Why non-statisticians and statisticians alike should prefer a diff-log over a percent change transformation for balance variables in PPNR stress testing.
Eight critical immediate steps toward a successful 2017 CCAR submission with the cycle shortened from 15 to 12 months.
Regression-based approach to deposit rate modeling that combines replicating portfolio concepts with econometric rigor.
Documentation of a bug in the SAS PROC AUTOREG procedure where models with no serial correlation fail the Godfrey serial correlation test.
Framework for building scalable CCAR model factories with emphasis on project frameworks, overcoming data sterility, and testing alternative approaches.