ZERO-TRUST AGENTS6 min read
Teaching an Agent Fleet to Distrust Documents: Zero-Trust M&A Diligence on Google Cloud
How Diligence Room uses isolated agents, hostile-document screening, and evidence gates to make autonomous deal diligence defensible.
Thinking insystems.
Divyansh
Agrawal
ML RESEARCH & ENGINEERING
I build and evaluate learning systems.
INFRASTRUCTURE / RL / EVALUATIONSWhat I'm working on ↗
Notebook
My writing ↗ATTENTION LABINTERACTIVE GUIDES
How attention
works.↗
Explore eight attention mechanisms through diagrams and experiments.
RECENT POSTSRSS ↗
SPECULATIVE DECODING6 min read
The Tiny Draft Model Hidden Inside Qwen
Qwen ships a one-layer multi-token predictor that vLLM and SGLang can use for native speculative decoding.
MULTI-AGENT SYSTEMS15 min read
Do LLM Agent Societies Adapt Their Values, or Eventually Die by Them?
What an evolutionary multi-agent simulation reveals about cultural persistence, selection, migration, and the danger of values that never bend.
My work
RESEARCH & TOOLSEVALUATION↗
WebBench
I test whether agents can distinguish repeated claims from independent evidence.
EXPLORE PROJECT ↗INFRASTRUCTURE↗
Hyperion
I build low-latency routing, caching, and observability for LLMs.
EXPLORE PROJECT ↗RL ENVIRONMENTS↗
FactorioRL
An agent-native Factorio environment for testing LLM agents, learned policies, and hybrid control.
EXPLORE PROJECT ↗ABOUT MEI'm still figuring out
I'm still figuring out
how models learn.
I move between infrastructure, reinforcement learning, and evaluations to build small systems that make model behaviour easier to test and understand.
Outside work, I love movies, travelling, photography, and the occasional hackathon.
More about me ↗