Researching a self-creating, continuously evolving agent harness
that can propose and test improvements to its own prompts,
policies, routing, tools, memory, and code.
For my mathematics research paper, I am using decision and
probability theory to study this as a sequential decision problem
under uncertainty. Which change should an agent try, what evidence
makes it credible, and when does its expected utility justify
promotion? The current Rust prototype provides the testing ground
for replayable traces, various tools, explicit permissions,
reproducible evaluations, and gated promotion with rollback. This
is currently work in progress.
A GPU inference research platform for Qwen3-family models,
combining validated serving paths, reproducible benchmarks, and
custom CUDA prototypes.
I use it to evaluate Qwen3.5 0.8B across Transformers and vLLM,
measuring latency, throughput, memory, token parity, continuous
batching, and multi-agent load. It also includes a custom Qwen3
decode-attention kernel and an experimental exact-window plus
compressed-memory path for long contexts.
A Python-supported macro research report on how industrial policy,
strategic controls, and national bottlenecks intersect with
artificial intelligence.
I combined quantitative research with written analysis to frame
the central thesis, develop forecastable claims, and connect AI
progress with the policy and infrastructure surrounding it.
An interactive Next.js application for practicing technical
analyst interviews and reviewing the concepts behind each
question.
I built the practice and review experience end to end, focusing on
a clear question flow and an easy way to revisit core concepts
while preparing for interviews.
