Lab
My research, explained in depth. My peer-reviewed work tackles a core weakness of deep learning models, the errors that build up as they predict further ahead, and my latest work is on how AI systems retrieve knowledge. Here I explain that work for readers outside the field.
Publications
8 peer-reviewed papers and 1 preprint.
Semantic Compression Trees: Multi-resolution knowledge retrieval via hierarchical semantic residuals
A new way to index documents for AI retrieval, in which each level of a hierarchy stores only what it adds to the level above. It matches standard retrieval quality with 30% fewer context tokens and scales far better with collection size; the paper also reports plainly that its top-down search step does not help.
arXiv, 2026
RefreshNet: Learning multiscale dynamics through hierarchical refreshing
Forecasts that step forward one small interval at a time build up error over long horizons. RefreshNet stacks models that work at different time scales and lets the slower ones periodically correct the faster ones.
Nonlinear Dynamics, 2024
Deep learning for self-tuning of control systems
Using deep learning to tune a control system's settings automatically instead of by hand.
ICETCI 2021, IEEE, 2021 · Best Paper Award
Topics
The areas I work and write in.
AI search and retrieval
Search and retrieval systems over large, messy data.
Enterprise AI agents
Agents in production: evaluation, failure modes, governance.
AI infrastructure economics
Inference cost, cloud cost, build versus buy.
AI security and risk
How AI systems fail and how they are attacked.
Engineering leadership
Small senior teams, architecture ownership, speed.
Scientific machine learning
Deep learning for forecasting dynamical systems.