Maxwell Zhang
Research, ideas & notes

Maxwell Zhang

Notes on agents, learning,
and evaluation.

A place to share research, experiments, and personal perspectives on building better AI agents.

Writing

01 / FIELD NOTES
WORKING DRAFT

Practical RSI for Enterprise Agents

An introduction to RSI, a changing enterprise environment, and the memory architecture we designed for ARISE—with two animated diagrams.

SELF-IMPROVEMENTAGENT MEMORYARISE
↗
WORKING DRAFT

Enterprise Agentic Evaluation

Why enterprise evaluation needs its own tasks, why gold data is expensive, and how DI-Bench helps—with an interactive walkthrough.

LLM AGENTSEVALUATIONDI-BENCH
↗

Selected research

02 / PUBLICATIONS

DI-Bench: Systematically generating in-domain benchmarks for data intelligence agents

Maxwell Zhang and collaborators

Generating domain-specific evaluation tasks that bring document knowledge and structured data together.

Paper & full author list ↗

A little about me

I work on LLMs and applied machine learning at Amazon. I am interested in building agents we can understand, evaluate, and trust, and in connecting research ideas with working systems.

I write to share research, technical notes, and ideas in progress. This is a place to think out loud, connect with other researchers and builders, and learn through discussion.

Agent Learning & Self-Improvement
How agents learn from experience and feedback, adapt their memory, and become more capable and efficient over time.
LLM Agents & Evaluation
Reasoning, tool use, enterprise data intelligence, and evaluations that reveal how and why agents fail.
AI Safety & Red Teaming
Exploring robustness, failure modes, and how to test the boundaries of agent behavior.
Multimodal AI
Exploring how agents understand and interact through images, video, and voice.